blob: fcf7265ca46fd84a65b647619f4e07ba996d7a9f [file] [log] [blame]
David Woodhouse53b381b2013-01-29 18:40:14 -05001/*
2 * Copyright (C) 2012 Fusion-io All rights reserved.
3 * Copyright (C) 2012 Intel Corp. All rights reserved.
4 *
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public
7 * License v2 as published by the Free Software Foundation.
8 *
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
12 * General Public License for more details.
13 *
14 * You should have received a copy of the GNU General Public
15 * License along with this program; if not, write to the
16 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
17 * Boston, MA 021110-1307, USA.
18 */
19#include <linux/sched.h>
20#include <linux/wait.h>
21#include <linux/bio.h>
22#include <linux/slab.h>
23#include <linux/buffer_head.h>
24#include <linux/blkdev.h>
25#include <linux/random.h>
26#include <linux/iocontext.h>
27#include <linux/capability.h>
28#include <linux/ratelimit.h>
29#include <linux/kthread.h>
30#include <linux/raid/pq.h>
31#include <linux/hash.h>
32#include <linux/list_sort.h>
33#include <linux/raid/xor.h>
Geert Uytterhoevend7011f52013-03-03 04:44:41 -070034#include <linux/vmalloc.h>
David Woodhouse53b381b2013-01-29 18:40:14 -050035#include <asm/div64.h>
David Woodhouse53b381b2013-01-29 18:40:14 -050036#include "ctree.h"
37#include "extent_map.h"
38#include "disk-io.h"
39#include "transaction.h"
40#include "print-tree.h"
41#include "volumes.h"
42#include "raid56.h"
43#include "async-thread.h"
44#include "check-integrity.h"
45#include "rcu-string.h"
46
47/* set when additional merges to this rbio are not allowed */
48#define RBIO_RMW_LOCKED_BIT 1
49
Chris Mason4ae10b32013-01-31 14:42:09 -050050/*
51 * set when this rbio is sitting in the hash, but it is just a cache
52 * of past RMW
53 */
54#define RBIO_CACHE_BIT 2
55
56/*
57 * set when it is safe to trust the stripe_pages for caching
58 */
59#define RBIO_CACHE_READY_BIT 3
60
Chris Mason4ae10b32013-01-31 14:42:09 -050061#define RBIO_CACHE_SIZE 1024
62
Miao Xie1b94b552014-11-06 16:14:21 +080063enum btrfs_rbio_ops {
Omar Sandovalb4ee1782015-06-19 11:52:50 -070064 BTRFS_RBIO_WRITE,
65 BTRFS_RBIO_READ_REBUILD,
66 BTRFS_RBIO_PARITY_SCRUB,
67 BTRFS_RBIO_REBUILD_MISSING,
Miao Xie1b94b552014-11-06 16:14:21 +080068};
69
David Woodhouse53b381b2013-01-29 18:40:14 -050070struct btrfs_raid_bio {
71 struct btrfs_fs_info *fs_info;
72 struct btrfs_bio *bbio;
73
David Woodhouse53b381b2013-01-29 18:40:14 -050074 /* while we're doing rmw on a stripe
75 * we put it into a hash table so we can
76 * lock the stripe and merge more rbios
77 * into it.
78 */
79 struct list_head hash_list;
80
81 /*
Chris Mason4ae10b32013-01-31 14:42:09 -050082 * LRU list for the stripe cache
83 */
84 struct list_head stripe_cache;
85
86 /*
David Woodhouse53b381b2013-01-29 18:40:14 -050087 * for scheduling work in the helper threads
88 */
89 struct btrfs_work work;
90
91 /*
92 * bio list and bio_list_lock are used
93 * to add more bios into the stripe
94 * in hopes of avoiding the full rmw
95 */
96 struct bio_list bio_list;
97 spinlock_t bio_list_lock;
98
Chris Mason6ac0f482013-01-31 14:42:28 -050099 /* also protected by the bio_list_lock, the
100 * plug list is used by the plugging code
101 * to collect partial bios while plugged. The
102 * stripe locking code also uses it to hand off
David Woodhouse53b381b2013-01-29 18:40:14 -0500103 * the stripe lock to the next pending IO
104 */
105 struct list_head plug_list;
106
107 /*
108 * flags that tell us if it is safe to
109 * merge with this bio
110 */
111 unsigned long flags;
112
113 /* size of each individual stripe on disk */
114 int stripe_len;
115
116 /* number of data stripes (no p/q) */
117 int nr_data;
118
Miao Xie2c8cdd62014-11-14 16:06:25 +0800119 int real_stripes;
120
Miao Xie5a6ac9e2014-11-06 17:20:58 +0800121 int stripe_npages;
David Woodhouse53b381b2013-01-29 18:40:14 -0500122 /*
123 * set if we're doing a parity rebuild
124 * for a read from higher up, which is handled
125 * differently from a parity rebuild as part of
126 * rmw
127 */
Miao Xie1b94b552014-11-06 16:14:21 +0800128 enum btrfs_rbio_ops operation;
David Woodhouse53b381b2013-01-29 18:40:14 -0500129
130 /* first bad stripe */
131 int faila;
132
133 /* second bad stripe (for raid6 use) */
134 int failb;
135
Miao Xie5a6ac9e2014-11-06 17:20:58 +0800136 int scrubp;
David Woodhouse53b381b2013-01-29 18:40:14 -0500137 /*
138 * number of pages needed to represent the full
139 * stripe
140 */
141 int nr_pages;
142
143 /*
144 * size of all the bios in the bio_list. This
145 * helps us decide if the rbio maps to a full
146 * stripe or not
147 */
148 int bio_list_bytes;
149
Miao Xie42452152014-11-25 16:39:28 +0800150 int generic_bio_cnt;
151
David Woodhouse53b381b2013-01-29 18:40:14 -0500152 atomic_t refs;
153
Miao Xieb89e1b02014-10-15 11:18:44 +0800154 atomic_t stripes_pending;
155
156 atomic_t error;
David Woodhouse53b381b2013-01-29 18:40:14 -0500157 /*
158 * these are two arrays of pointers. We allocate the
159 * rbio big enough to hold them both and setup their
160 * locations when the rbio is allocated
161 */
162
163 /* pointers to pages that we allocated for
164 * reading/writing stripes directly from the disk (including P/Q)
165 */
166 struct page **stripe_pages;
167
168 /*
169 * pointers to the pages in the bio_list. Stored
170 * here for faster lookup
171 */
172 struct page **bio_pages;
Miao Xie5a6ac9e2014-11-06 17:20:58 +0800173
174 /*
175 * bitmap to record which horizontal stripe has data
176 */
177 unsigned long *dbitmap;
David Woodhouse53b381b2013-01-29 18:40:14 -0500178};
179
180static int __raid56_parity_recover(struct btrfs_raid_bio *rbio);
181static noinline void finish_rmw(struct btrfs_raid_bio *rbio);
182static void rmw_work(struct btrfs_work *work);
183static void read_rebuild_work(struct btrfs_work *work);
184static void async_rmw_stripe(struct btrfs_raid_bio *rbio);
185static void async_read_rebuild(struct btrfs_raid_bio *rbio);
186static int fail_bio_stripe(struct btrfs_raid_bio *rbio, struct bio *bio);
187static int fail_rbio_index(struct btrfs_raid_bio *rbio, int failed);
188static void __free_raid_bio(struct btrfs_raid_bio *rbio);
189static void index_rbio_pages(struct btrfs_raid_bio *rbio);
190static int alloc_rbio_pages(struct btrfs_raid_bio *rbio);
191
Miao Xie5a6ac9e2014-11-06 17:20:58 +0800192static noinline void finish_parity_scrub(struct btrfs_raid_bio *rbio,
193 int need_check);
194static void async_scrub_parity(struct btrfs_raid_bio *rbio);
195
David Woodhouse53b381b2013-01-29 18:40:14 -0500196/*
197 * the stripe hash table is used for locking, and to collect
198 * bios in hopes of making a full stripe
199 */
200int btrfs_alloc_stripe_hash_table(struct btrfs_fs_info *info)
201{
202 struct btrfs_stripe_hash_table *table;
203 struct btrfs_stripe_hash_table *x;
204 struct btrfs_stripe_hash *cur;
205 struct btrfs_stripe_hash *h;
206 int num_entries = 1 << BTRFS_STRIPE_HASH_TABLE_BITS;
207 int i;
David Sterba83c82662013-03-01 15:03:00 +0000208 int table_size;
David Woodhouse53b381b2013-01-29 18:40:14 -0500209
210 if (info->stripe_hash_table)
211 return 0;
212
David Sterba83c82662013-03-01 15:03:00 +0000213 /*
214 * The table is large, starting with order 4 and can go as high as
215 * order 7 in case lock debugging is turned on.
216 *
217 * Try harder to allocate and fallback to vmalloc to lower the chance
218 * of a failing mount.
219 */
220 table_size = sizeof(*table) + sizeof(*h) * num_entries;
221 table = kzalloc(table_size, GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT);
222 if (!table) {
223 table = vzalloc(table_size);
224 if (!table)
225 return -ENOMEM;
226 }
David Woodhouse53b381b2013-01-29 18:40:14 -0500227
Chris Mason4ae10b32013-01-31 14:42:09 -0500228 spin_lock_init(&table->cache_lock);
229 INIT_LIST_HEAD(&table->stripe_cache);
230
David Woodhouse53b381b2013-01-29 18:40:14 -0500231 h = table->table;
232
233 for (i = 0; i < num_entries; i++) {
234 cur = h + i;
235 INIT_LIST_HEAD(&cur->hash_list);
236 spin_lock_init(&cur->lock);
237 init_waitqueue_head(&cur->wait);
238 }
239
240 x = cmpxchg(&info->stripe_hash_table, NULL, table);
Wang Shilongf7493032014-11-22 21:13:10 +0800241 if (x)
242 kvfree(x);
David Woodhouse53b381b2013-01-29 18:40:14 -0500243 return 0;
244}
245
246/*
Chris Mason4ae10b32013-01-31 14:42:09 -0500247 * caching an rbio means to copy anything from the
248 * bio_pages array into the stripe_pages array. We
249 * use the page uptodate bit in the stripe cache array
250 * to indicate if it has valid data
251 *
252 * once the caching is done, we set the cache ready
253 * bit.
254 */
255static void cache_rbio_pages(struct btrfs_raid_bio *rbio)
256{
257 int i;
258 char *s;
259 char *d;
260 int ret;
261
262 ret = alloc_rbio_pages(rbio);
263 if (ret)
264 return;
265
266 for (i = 0; i < rbio->nr_pages; i++) {
267 if (!rbio->bio_pages[i])
268 continue;
269
270 s = kmap(rbio->bio_pages[i]);
271 d = kmap(rbio->stripe_pages[i]);
272
273 memcpy(d, s, PAGE_CACHE_SIZE);
274
275 kunmap(rbio->bio_pages[i]);
276 kunmap(rbio->stripe_pages[i]);
277 SetPageUptodate(rbio->stripe_pages[i]);
278 }
279 set_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
280}
281
282/*
David Woodhouse53b381b2013-01-29 18:40:14 -0500283 * we hash on the first logical address of the stripe
284 */
285static int rbio_bucket(struct btrfs_raid_bio *rbio)
286{
Zhao Lei8e5cfb52015-01-20 15:11:33 +0800287 u64 num = rbio->bbio->raid_map[0];
David Woodhouse53b381b2013-01-29 18:40:14 -0500288
289 /*
290 * we shift down quite a bit. We're using byte
291 * addressing, and most of the lower bits are zeros.
292 * This tends to upset hash_64, and it consistently
293 * returns just one or two different values.
294 *
295 * shifting off the lower bits fixes things.
296 */
297 return hash_64(num >> 16, BTRFS_STRIPE_HASH_TABLE_BITS);
298}
299
300/*
Chris Mason4ae10b32013-01-31 14:42:09 -0500301 * stealing an rbio means taking all the uptodate pages from the stripe
302 * array in the source rbio and putting them into the destination rbio
303 */
304static void steal_rbio(struct btrfs_raid_bio *src, struct btrfs_raid_bio *dest)
305{
306 int i;
307 struct page *s;
308 struct page *d;
309
310 if (!test_bit(RBIO_CACHE_READY_BIT, &src->flags))
311 return;
312
313 for (i = 0; i < dest->nr_pages; i++) {
314 s = src->stripe_pages[i];
315 if (!s || !PageUptodate(s)) {
316 continue;
317 }
318
319 d = dest->stripe_pages[i];
320 if (d)
321 __free_page(d);
322
323 dest->stripe_pages[i] = s;
324 src->stripe_pages[i] = NULL;
325 }
326}
327
328/*
David Woodhouse53b381b2013-01-29 18:40:14 -0500329 * merging means we take the bio_list from the victim and
330 * splice it into the destination. The victim should
331 * be discarded afterwards.
332 *
333 * must be called with dest->rbio_list_lock held
334 */
335static void merge_rbio(struct btrfs_raid_bio *dest,
336 struct btrfs_raid_bio *victim)
337{
338 bio_list_merge(&dest->bio_list, &victim->bio_list);
339 dest->bio_list_bytes += victim->bio_list_bytes;
Miao Xie42452152014-11-25 16:39:28 +0800340 dest->generic_bio_cnt += victim->generic_bio_cnt;
David Woodhouse53b381b2013-01-29 18:40:14 -0500341 bio_list_init(&victim->bio_list);
342}
343
344/*
Chris Mason4ae10b32013-01-31 14:42:09 -0500345 * used to prune items that are in the cache. The caller
346 * must hold the hash table lock.
347 */
348static void __remove_rbio_from_cache(struct btrfs_raid_bio *rbio)
349{
350 int bucket = rbio_bucket(rbio);
351 struct btrfs_stripe_hash_table *table;
352 struct btrfs_stripe_hash *h;
353 int freeit = 0;
354
355 /*
356 * check the bit again under the hash table lock.
357 */
358 if (!test_bit(RBIO_CACHE_BIT, &rbio->flags))
359 return;
360
361 table = rbio->fs_info->stripe_hash_table;
362 h = table->table + bucket;
363
364 /* hold the lock for the bucket because we may be
365 * removing it from the hash table
366 */
367 spin_lock(&h->lock);
368
369 /*
370 * hold the lock for the bio list because we need
371 * to make sure the bio list is empty
372 */
373 spin_lock(&rbio->bio_list_lock);
374
375 if (test_and_clear_bit(RBIO_CACHE_BIT, &rbio->flags)) {
376 list_del_init(&rbio->stripe_cache);
377 table->cache_size -= 1;
378 freeit = 1;
379
380 /* if the bio list isn't empty, this rbio is
381 * still involved in an IO. We take it out
382 * of the cache list, and drop the ref that
383 * was held for the list.
384 *
385 * If the bio_list was empty, we also remove
386 * the rbio from the hash_table, and drop
387 * the corresponding ref
388 */
389 if (bio_list_empty(&rbio->bio_list)) {
390 if (!list_empty(&rbio->hash_list)) {
391 list_del_init(&rbio->hash_list);
392 atomic_dec(&rbio->refs);
393 BUG_ON(!list_empty(&rbio->plug_list));
394 }
395 }
396 }
397
398 spin_unlock(&rbio->bio_list_lock);
399 spin_unlock(&h->lock);
400
401 if (freeit)
402 __free_raid_bio(rbio);
403}
404
405/*
406 * prune a given rbio from the cache
407 */
408static void remove_rbio_from_cache(struct btrfs_raid_bio *rbio)
409{
410 struct btrfs_stripe_hash_table *table;
411 unsigned long flags;
412
413 if (!test_bit(RBIO_CACHE_BIT, &rbio->flags))
414 return;
415
416 table = rbio->fs_info->stripe_hash_table;
417
418 spin_lock_irqsave(&table->cache_lock, flags);
419 __remove_rbio_from_cache(rbio);
420 spin_unlock_irqrestore(&table->cache_lock, flags);
421}
422
423/*
424 * remove everything in the cache
425 */
Eric Sandeen48a3b632013-04-25 20:41:01 +0000426static void btrfs_clear_rbio_cache(struct btrfs_fs_info *info)
Chris Mason4ae10b32013-01-31 14:42:09 -0500427{
428 struct btrfs_stripe_hash_table *table;
429 unsigned long flags;
430 struct btrfs_raid_bio *rbio;
431
432 table = info->stripe_hash_table;
433
434 spin_lock_irqsave(&table->cache_lock, flags);
435 while (!list_empty(&table->stripe_cache)) {
436 rbio = list_entry(table->stripe_cache.next,
437 struct btrfs_raid_bio,
438 stripe_cache);
439 __remove_rbio_from_cache(rbio);
440 }
441 spin_unlock_irqrestore(&table->cache_lock, flags);
442}
443
444/*
445 * remove all cached entries and free the hash table
446 * used by unmount
David Woodhouse53b381b2013-01-29 18:40:14 -0500447 */
448void btrfs_free_stripe_hash_table(struct btrfs_fs_info *info)
449{
450 if (!info->stripe_hash_table)
451 return;
Chris Mason4ae10b32013-01-31 14:42:09 -0500452 btrfs_clear_rbio_cache(info);
Wang Shilongf7493032014-11-22 21:13:10 +0800453 kvfree(info->stripe_hash_table);
David Woodhouse53b381b2013-01-29 18:40:14 -0500454 info->stripe_hash_table = NULL;
455}
456
457/*
Chris Mason4ae10b32013-01-31 14:42:09 -0500458 * insert an rbio into the stripe cache. It
459 * must have already been prepared by calling
460 * cache_rbio_pages
461 *
462 * If this rbio was already cached, it gets
463 * moved to the front of the lru.
464 *
465 * If the size of the rbio cache is too big, we
466 * prune an item.
467 */
468static void cache_rbio(struct btrfs_raid_bio *rbio)
469{
470 struct btrfs_stripe_hash_table *table;
471 unsigned long flags;
472
473 if (!test_bit(RBIO_CACHE_READY_BIT, &rbio->flags))
474 return;
475
476 table = rbio->fs_info->stripe_hash_table;
477
478 spin_lock_irqsave(&table->cache_lock, flags);
479 spin_lock(&rbio->bio_list_lock);
480
481 /* bump our ref if we were not in the list before */
482 if (!test_and_set_bit(RBIO_CACHE_BIT, &rbio->flags))
483 atomic_inc(&rbio->refs);
484
485 if (!list_empty(&rbio->stripe_cache)){
486 list_move(&rbio->stripe_cache, &table->stripe_cache);
487 } else {
488 list_add(&rbio->stripe_cache, &table->stripe_cache);
489 table->cache_size += 1;
490 }
491
492 spin_unlock(&rbio->bio_list_lock);
493
494 if (table->cache_size > RBIO_CACHE_SIZE) {
495 struct btrfs_raid_bio *found;
496
497 found = list_entry(table->stripe_cache.prev,
498 struct btrfs_raid_bio,
499 stripe_cache);
500
501 if (found != rbio)
502 __remove_rbio_from_cache(found);
503 }
504
505 spin_unlock_irqrestore(&table->cache_lock, flags);
506 return;
507}
508
509/*
David Woodhouse53b381b2013-01-29 18:40:14 -0500510 * helper function to run the xor_blocks api. It is only
511 * able to do MAX_XOR_BLOCKS at a time, so we need to
512 * loop through.
513 */
514static void run_xor(void **pages, int src_cnt, ssize_t len)
515{
516 int src_off = 0;
517 int xor_src_cnt = 0;
518 void *dest = pages[src_cnt];
519
520 while(src_cnt > 0) {
521 xor_src_cnt = min(src_cnt, MAX_XOR_BLOCKS);
522 xor_blocks(xor_src_cnt, len, dest, pages + src_off);
523
524 src_cnt -= xor_src_cnt;
525 src_off += xor_src_cnt;
526 }
527}
528
529/*
530 * returns true if the bio list inside this rbio
531 * covers an entire stripe (no rmw required).
532 * Must be called with the bio list lock held, or
533 * at a time when you know it is impossible to add
534 * new bios into the list
535 */
536static int __rbio_is_full(struct btrfs_raid_bio *rbio)
537{
538 unsigned long size = rbio->bio_list_bytes;
539 int ret = 1;
540
541 if (size != rbio->nr_data * rbio->stripe_len)
542 ret = 0;
543
544 BUG_ON(size > rbio->nr_data * rbio->stripe_len);
545 return ret;
546}
547
548static int rbio_is_full(struct btrfs_raid_bio *rbio)
549{
550 unsigned long flags;
551 int ret;
552
553 spin_lock_irqsave(&rbio->bio_list_lock, flags);
554 ret = __rbio_is_full(rbio);
555 spin_unlock_irqrestore(&rbio->bio_list_lock, flags);
556 return ret;
557}
558
559/*
560 * returns 1 if it is safe to merge two rbios together.
561 * The merging is safe if the two rbios correspond to
562 * the same stripe and if they are both going in the same
563 * direction (read vs write), and if neither one is
564 * locked for final IO
565 *
566 * The caller is responsible for locking such that
567 * rmw_locked is safe to test
568 */
569static int rbio_can_merge(struct btrfs_raid_bio *last,
570 struct btrfs_raid_bio *cur)
571{
572 if (test_bit(RBIO_RMW_LOCKED_BIT, &last->flags) ||
573 test_bit(RBIO_RMW_LOCKED_BIT, &cur->flags))
574 return 0;
575
Chris Mason4ae10b32013-01-31 14:42:09 -0500576 /*
577 * we can't merge with cached rbios, since the
578 * idea is that when we merge the destination
579 * rbio is going to run our IO for us. We can
580 * steal from cached rbio's though, other functions
581 * handle that.
582 */
583 if (test_bit(RBIO_CACHE_BIT, &last->flags) ||
584 test_bit(RBIO_CACHE_BIT, &cur->flags))
585 return 0;
586
Zhao Lei8e5cfb52015-01-20 15:11:33 +0800587 if (last->bbio->raid_map[0] !=
588 cur->bbio->raid_map[0])
David Woodhouse53b381b2013-01-29 18:40:14 -0500589 return 0;
590
Miao Xie5a6ac9e2014-11-06 17:20:58 +0800591 /* we can't merge with different operations */
592 if (last->operation != cur->operation)
David Woodhouse53b381b2013-01-29 18:40:14 -0500593 return 0;
Miao Xie5a6ac9e2014-11-06 17:20:58 +0800594 /*
595 * We've need read the full stripe from the drive.
596 * check and repair the parity and write the new results.
597 *
598 * We're not allowed to add any new bios to the
599 * bio list here, anyone else that wants to
600 * change this stripe needs to do their own rmw.
601 */
602 if (last->operation == BTRFS_RBIO_PARITY_SCRUB ||
603 cur->operation == BTRFS_RBIO_PARITY_SCRUB)
604 return 0;
David Woodhouse53b381b2013-01-29 18:40:14 -0500605
Omar Sandovalb4ee1782015-06-19 11:52:50 -0700606 if (last->operation == BTRFS_RBIO_REBUILD_MISSING ||
607 cur->operation == BTRFS_RBIO_REBUILD_MISSING)
608 return 0;
609
David Woodhouse53b381b2013-01-29 18:40:14 -0500610 return 1;
611}
612
613/*
614 * helper to index into the pstripe
615 */
616static struct page *rbio_pstripe_page(struct btrfs_raid_bio *rbio, int index)
617{
618 index += (rbio->nr_data * rbio->stripe_len) >> PAGE_CACHE_SHIFT;
619 return rbio->stripe_pages[index];
620}
621
622/*
623 * helper to index into the qstripe, returns null
624 * if there is no qstripe
625 */
626static struct page *rbio_qstripe_page(struct btrfs_raid_bio *rbio, int index)
627{
Miao Xie2c8cdd62014-11-14 16:06:25 +0800628 if (rbio->nr_data + 1 == rbio->real_stripes)
David Woodhouse53b381b2013-01-29 18:40:14 -0500629 return NULL;
630
631 index += ((rbio->nr_data + 1) * rbio->stripe_len) >>
632 PAGE_CACHE_SHIFT;
633 return rbio->stripe_pages[index];
634}
635
636/*
637 * The first stripe in the table for a logical address
638 * has the lock. rbios are added in one of three ways:
639 *
640 * 1) Nobody has the stripe locked yet. The rbio is given
641 * the lock and 0 is returned. The caller must start the IO
642 * themselves.
643 *
644 * 2) Someone has the stripe locked, but we're able to merge
645 * with the lock owner. The rbio is freed and the IO will
646 * start automatically along with the existing rbio. 1 is returned.
647 *
648 * 3) Someone has the stripe locked, but we're not able to merge.
649 * The rbio is added to the lock owner's plug list, or merged into
650 * an rbio already on the plug list. When the lock owner unlocks,
651 * the next rbio on the list is run and the IO is started automatically.
652 * 1 is returned
653 *
654 * If we return 0, the caller still owns the rbio and must continue with
655 * IO submission. If we return 1, the caller must assume the rbio has
656 * already been freed.
657 */
658static noinline int lock_stripe_add(struct btrfs_raid_bio *rbio)
659{
660 int bucket = rbio_bucket(rbio);
661 struct btrfs_stripe_hash *h = rbio->fs_info->stripe_hash_table->table + bucket;
662 struct btrfs_raid_bio *cur;
663 struct btrfs_raid_bio *pending;
664 unsigned long flags;
665 DEFINE_WAIT(wait);
666 struct btrfs_raid_bio *freeit = NULL;
Chris Mason4ae10b32013-01-31 14:42:09 -0500667 struct btrfs_raid_bio *cache_drop = NULL;
David Woodhouse53b381b2013-01-29 18:40:14 -0500668 int ret = 0;
669 int walk = 0;
670
671 spin_lock_irqsave(&h->lock, flags);
672 list_for_each_entry(cur, &h->hash_list, hash_list) {
673 walk++;
Zhao Lei8e5cfb52015-01-20 15:11:33 +0800674 if (cur->bbio->raid_map[0] == rbio->bbio->raid_map[0]) {
David Woodhouse53b381b2013-01-29 18:40:14 -0500675 spin_lock(&cur->bio_list_lock);
676
Chris Mason4ae10b32013-01-31 14:42:09 -0500677 /* can we steal this cached rbio's pages? */
678 if (bio_list_empty(&cur->bio_list) &&
679 list_empty(&cur->plug_list) &&
680 test_bit(RBIO_CACHE_BIT, &cur->flags) &&
681 !test_bit(RBIO_RMW_LOCKED_BIT, &cur->flags)) {
682 list_del_init(&cur->hash_list);
683 atomic_dec(&cur->refs);
684
685 steal_rbio(cur, rbio);
686 cache_drop = cur;
687 spin_unlock(&cur->bio_list_lock);
688
689 goto lockit;
690 }
691
David Woodhouse53b381b2013-01-29 18:40:14 -0500692 /* can we merge into the lock owner? */
693 if (rbio_can_merge(cur, rbio)) {
694 merge_rbio(cur, rbio);
695 spin_unlock(&cur->bio_list_lock);
696 freeit = rbio;
697 ret = 1;
698 goto out;
699 }
700
Chris Mason4ae10b32013-01-31 14:42:09 -0500701
David Woodhouse53b381b2013-01-29 18:40:14 -0500702 /*
703 * we couldn't merge with the running
704 * rbio, see if we can merge with the
705 * pending ones. We don't have to
706 * check for rmw_locked because there
707 * is no way they are inside finish_rmw
708 * right now
709 */
710 list_for_each_entry(pending, &cur->plug_list,
711 plug_list) {
712 if (rbio_can_merge(pending, rbio)) {
713 merge_rbio(pending, rbio);
714 spin_unlock(&cur->bio_list_lock);
715 freeit = rbio;
716 ret = 1;
717 goto out;
718 }
719 }
720
721 /* no merging, put us on the tail of the plug list,
722 * our rbio will be started with the currently
723 * running rbio unlocks
724 */
725 list_add_tail(&rbio->plug_list, &cur->plug_list);
726 spin_unlock(&cur->bio_list_lock);
727 ret = 1;
728 goto out;
729 }
730 }
Chris Mason4ae10b32013-01-31 14:42:09 -0500731lockit:
David Woodhouse53b381b2013-01-29 18:40:14 -0500732 atomic_inc(&rbio->refs);
733 list_add(&rbio->hash_list, &h->hash_list);
734out:
735 spin_unlock_irqrestore(&h->lock, flags);
Chris Mason4ae10b32013-01-31 14:42:09 -0500736 if (cache_drop)
737 remove_rbio_from_cache(cache_drop);
David Woodhouse53b381b2013-01-29 18:40:14 -0500738 if (freeit)
739 __free_raid_bio(freeit);
740 return ret;
741}
742
743/*
744 * called as rmw or parity rebuild is completed. If the plug list has more
745 * rbios waiting for this stripe, the next one on the list will be started
746 */
747static noinline void unlock_stripe(struct btrfs_raid_bio *rbio)
748{
749 int bucket;
750 struct btrfs_stripe_hash *h;
751 unsigned long flags;
Chris Mason4ae10b32013-01-31 14:42:09 -0500752 int keep_cache = 0;
David Woodhouse53b381b2013-01-29 18:40:14 -0500753
754 bucket = rbio_bucket(rbio);
755 h = rbio->fs_info->stripe_hash_table->table + bucket;
756
Chris Mason4ae10b32013-01-31 14:42:09 -0500757 if (list_empty(&rbio->plug_list))
758 cache_rbio(rbio);
759
David Woodhouse53b381b2013-01-29 18:40:14 -0500760 spin_lock_irqsave(&h->lock, flags);
761 spin_lock(&rbio->bio_list_lock);
762
763 if (!list_empty(&rbio->hash_list)) {
Chris Mason4ae10b32013-01-31 14:42:09 -0500764 /*
765 * if we're still cached and there is no other IO
766 * to perform, just leave this rbio here for others
767 * to steal from later
768 */
769 if (list_empty(&rbio->plug_list) &&
770 test_bit(RBIO_CACHE_BIT, &rbio->flags)) {
771 keep_cache = 1;
772 clear_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
773 BUG_ON(!bio_list_empty(&rbio->bio_list));
774 goto done;
775 }
David Woodhouse53b381b2013-01-29 18:40:14 -0500776
777 list_del_init(&rbio->hash_list);
778 atomic_dec(&rbio->refs);
779
780 /*
781 * we use the plug list to hold all the rbios
782 * waiting for the chance to lock this stripe.
783 * hand the lock over to one of them.
784 */
785 if (!list_empty(&rbio->plug_list)) {
786 struct btrfs_raid_bio *next;
787 struct list_head *head = rbio->plug_list.next;
788
789 next = list_entry(head, struct btrfs_raid_bio,
790 plug_list);
791
792 list_del_init(&rbio->plug_list);
793
794 list_add(&next->hash_list, &h->hash_list);
795 atomic_inc(&next->refs);
796 spin_unlock(&rbio->bio_list_lock);
797 spin_unlock_irqrestore(&h->lock, flags);
798
Miao Xie1b94b552014-11-06 16:14:21 +0800799 if (next->operation == BTRFS_RBIO_READ_REBUILD)
David Woodhouse53b381b2013-01-29 18:40:14 -0500800 async_read_rebuild(next);
Omar Sandovalb4ee1782015-06-19 11:52:50 -0700801 else if (next->operation == BTRFS_RBIO_REBUILD_MISSING) {
802 steal_rbio(rbio, next);
803 async_read_rebuild(next);
804 } else if (next->operation == BTRFS_RBIO_WRITE) {
Chris Mason4ae10b32013-01-31 14:42:09 -0500805 steal_rbio(rbio, next);
David Woodhouse53b381b2013-01-29 18:40:14 -0500806 async_rmw_stripe(next);
Miao Xie5a6ac9e2014-11-06 17:20:58 +0800807 } else if (next->operation == BTRFS_RBIO_PARITY_SCRUB) {
808 steal_rbio(rbio, next);
809 async_scrub_parity(next);
Chris Mason4ae10b32013-01-31 14:42:09 -0500810 }
David Woodhouse53b381b2013-01-29 18:40:14 -0500811
812 goto done_nolock;
David Woodhouse53b381b2013-01-29 18:40:14 -0500813 } else if (waitqueue_active(&h->wait)) {
814 spin_unlock(&rbio->bio_list_lock);
815 spin_unlock_irqrestore(&h->lock, flags);
816 wake_up(&h->wait);
817 goto done_nolock;
818 }
819 }
Chris Mason4ae10b32013-01-31 14:42:09 -0500820done:
David Woodhouse53b381b2013-01-29 18:40:14 -0500821 spin_unlock(&rbio->bio_list_lock);
822 spin_unlock_irqrestore(&h->lock, flags);
823
824done_nolock:
Chris Mason4ae10b32013-01-31 14:42:09 -0500825 if (!keep_cache)
826 remove_rbio_from_cache(rbio);
David Woodhouse53b381b2013-01-29 18:40:14 -0500827}
828
829static void __free_raid_bio(struct btrfs_raid_bio *rbio)
830{
831 int i;
832
833 WARN_ON(atomic_read(&rbio->refs) < 0);
834 if (!atomic_dec_and_test(&rbio->refs))
835 return;
836
Chris Mason4ae10b32013-01-31 14:42:09 -0500837 WARN_ON(!list_empty(&rbio->stripe_cache));
David Woodhouse53b381b2013-01-29 18:40:14 -0500838 WARN_ON(!list_empty(&rbio->hash_list));
839 WARN_ON(!bio_list_empty(&rbio->bio_list));
840
841 for (i = 0; i < rbio->nr_pages; i++) {
842 if (rbio->stripe_pages[i]) {
843 __free_page(rbio->stripe_pages[i]);
844 rbio->stripe_pages[i] = NULL;
845 }
846 }
Miao Xieaf8e2d12014-10-23 14:42:50 +0800847
Zhao Lei6e9606d2015-01-20 15:11:34 +0800848 btrfs_put_bbio(rbio->bbio);
David Woodhouse53b381b2013-01-29 18:40:14 -0500849 kfree(rbio);
850}
851
852static void free_raid_bio(struct btrfs_raid_bio *rbio)
853{
854 unlock_stripe(rbio);
855 __free_raid_bio(rbio);
856}
857
858/*
859 * this frees the rbio and runs through all the bios in the
860 * bio_list and calls end_io on them
861 */
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200862static void rbio_orig_end_io(struct btrfs_raid_bio *rbio, int err)
David Woodhouse53b381b2013-01-29 18:40:14 -0500863{
864 struct bio *cur = bio_list_get(&rbio->bio_list);
865 struct bio *next;
Miao Xie42452152014-11-25 16:39:28 +0800866
867 if (rbio->generic_bio_cnt)
868 btrfs_bio_counter_sub(rbio->fs_info, rbio->generic_bio_cnt);
869
David Woodhouse53b381b2013-01-29 18:40:14 -0500870 free_raid_bio(rbio);
871
872 while (cur) {
873 next = cur->bi_next;
874 cur->bi_next = NULL;
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200875 cur->bi_error = err;
876 bio_endio(cur);
David Woodhouse53b381b2013-01-29 18:40:14 -0500877 cur = next;
878 }
879}
880
881/*
882 * end io function used by finish_rmw. When we finally
883 * get here, we've written a full stripe
884 */
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200885static void raid_write_end_io(struct bio *bio)
David Woodhouse53b381b2013-01-29 18:40:14 -0500886{
887 struct btrfs_raid_bio *rbio = bio->bi_private;
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200888 int err = bio->bi_error;
David Woodhouse53b381b2013-01-29 18:40:14 -0500889
890 if (err)
891 fail_bio_stripe(rbio, bio);
892
893 bio_put(bio);
894
Miao Xieb89e1b02014-10-15 11:18:44 +0800895 if (!atomic_dec_and_test(&rbio->stripes_pending))
David Woodhouse53b381b2013-01-29 18:40:14 -0500896 return;
897
898 err = 0;
899
900 /* OK, we have read all the stripes we need to. */
Miao Xieb89e1b02014-10-15 11:18:44 +0800901 if (atomic_read(&rbio->error) > rbio->bbio->max_errors)
David Woodhouse53b381b2013-01-29 18:40:14 -0500902 err = -EIO;
903
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200904 rbio_orig_end_io(rbio, err);
David Woodhouse53b381b2013-01-29 18:40:14 -0500905 return;
906}
907
908/*
909 * the read/modify/write code wants to use the original bio for
910 * any pages it included, and then use the rbio for everything
911 * else. This function decides if a given index (stripe number)
912 * and page number in that stripe fall inside the original bio
913 * or the rbio.
914 *
915 * if you set bio_list_only, you'll get a NULL back for any ranges
916 * that are outside the bio_list
917 *
918 * This doesn't take any refs on anything, you get a bare page pointer
919 * and the caller must bump refs as required.
920 *
921 * You must call index_rbio_pages once before you can trust
922 * the answers from this function.
923 */
924static struct page *page_in_rbio(struct btrfs_raid_bio *rbio,
925 int index, int pagenr, int bio_list_only)
926{
927 int chunk_page;
928 struct page *p = NULL;
929
930 chunk_page = index * (rbio->stripe_len >> PAGE_SHIFT) + pagenr;
931
932 spin_lock_irq(&rbio->bio_list_lock);
933 p = rbio->bio_pages[chunk_page];
934 spin_unlock_irq(&rbio->bio_list_lock);
935
936 if (p || bio_list_only)
937 return p;
938
939 return rbio->stripe_pages[chunk_page];
940}
941
942/*
943 * number of pages we need for the entire stripe across all the
944 * drives
945 */
946static unsigned long rbio_nr_pages(unsigned long stripe_len, int nr_stripes)
947{
948 unsigned long nr = stripe_len * nr_stripes;
David Sterbaed6078f2014-06-05 01:59:57 +0200949 return DIV_ROUND_UP(nr, PAGE_CACHE_SIZE);
David Woodhouse53b381b2013-01-29 18:40:14 -0500950}
951
952/*
953 * allocation and initial setup for the btrfs_raid_bio. Not
954 * this does not allocate any pages for rbio->pages.
955 */
956static struct btrfs_raid_bio *alloc_rbio(struct btrfs_root *root,
Zhao Lei8e5cfb52015-01-20 15:11:33 +0800957 struct btrfs_bio *bbio, u64 stripe_len)
David Woodhouse53b381b2013-01-29 18:40:14 -0500958{
959 struct btrfs_raid_bio *rbio;
960 int nr_data = 0;
Miao Xie2c8cdd62014-11-14 16:06:25 +0800961 int real_stripes = bbio->num_stripes - bbio->num_tgtdevs;
962 int num_pages = rbio_nr_pages(stripe_len, real_stripes);
Miao Xie5a6ac9e2014-11-06 17:20:58 +0800963 int stripe_npages = DIV_ROUND_UP(stripe_len, PAGE_SIZE);
David Woodhouse53b381b2013-01-29 18:40:14 -0500964 void *p;
965
Miao Xie5a6ac9e2014-11-06 17:20:58 +0800966 rbio = kzalloc(sizeof(*rbio) + num_pages * sizeof(struct page *) * 2 +
967 DIV_ROUND_UP(stripe_npages, BITS_PER_LONG / 8),
David Woodhouse53b381b2013-01-29 18:40:14 -0500968 GFP_NOFS);
Miao Xieaf8e2d12014-10-23 14:42:50 +0800969 if (!rbio)
David Woodhouse53b381b2013-01-29 18:40:14 -0500970 return ERR_PTR(-ENOMEM);
David Woodhouse53b381b2013-01-29 18:40:14 -0500971
972 bio_list_init(&rbio->bio_list);
973 INIT_LIST_HEAD(&rbio->plug_list);
974 spin_lock_init(&rbio->bio_list_lock);
Chris Mason4ae10b32013-01-31 14:42:09 -0500975 INIT_LIST_HEAD(&rbio->stripe_cache);
David Woodhouse53b381b2013-01-29 18:40:14 -0500976 INIT_LIST_HEAD(&rbio->hash_list);
977 rbio->bbio = bbio;
David Woodhouse53b381b2013-01-29 18:40:14 -0500978 rbio->fs_info = root->fs_info;
979 rbio->stripe_len = stripe_len;
980 rbio->nr_pages = num_pages;
Miao Xie2c8cdd62014-11-14 16:06:25 +0800981 rbio->real_stripes = real_stripes;
Miao Xie5a6ac9e2014-11-06 17:20:58 +0800982 rbio->stripe_npages = stripe_npages;
David Woodhouse53b381b2013-01-29 18:40:14 -0500983 rbio->faila = -1;
984 rbio->failb = -1;
985 atomic_set(&rbio->refs, 1);
Miao Xieb89e1b02014-10-15 11:18:44 +0800986 atomic_set(&rbio->error, 0);
987 atomic_set(&rbio->stripes_pending, 0);
David Woodhouse53b381b2013-01-29 18:40:14 -0500988
989 /*
990 * the stripe_pages and bio_pages array point to the extra
991 * memory we allocated past the end of the rbio
992 */
993 p = rbio + 1;
994 rbio->stripe_pages = p;
995 rbio->bio_pages = p + sizeof(struct page *) * num_pages;
Miao Xie5a6ac9e2014-11-06 17:20:58 +0800996 rbio->dbitmap = p + sizeof(struct page *) * num_pages * 2;
David Woodhouse53b381b2013-01-29 18:40:14 -0500997
Zhao Lei10f11902015-01-20 15:11:43 +0800998 if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID5)
999 nr_data = real_stripes - 1;
1000 else if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID6)
Miao Xie2c8cdd62014-11-14 16:06:25 +08001001 nr_data = real_stripes - 2;
David Woodhouse53b381b2013-01-29 18:40:14 -05001002 else
Zhao Lei10f11902015-01-20 15:11:43 +08001003 BUG();
David Woodhouse53b381b2013-01-29 18:40:14 -05001004
1005 rbio->nr_data = nr_data;
1006 return rbio;
1007}
1008
1009/* allocate pages for all the stripes in the bio, including parity */
1010static int alloc_rbio_pages(struct btrfs_raid_bio *rbio)
1011{
1012 int i;
1013 struct page *page;
1014
1015 for (i = 0; i < rbio->nr_pages; i++) {
1016 if (rbio->stripe_pages[i])
1017 continue;
1018 page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1019 if (!page)
1020 return -ENOMEM;
1021 rbio->stripe_pages[i] = page;
1022 ClearPageUptodate(page);
1023 }
1024 return 0;
1025}
1026
1027/* allocate pages for just the p/q stripes */
1028static int alloc_rbio_parity_pages(struct btrfs_raid_bio *rbio)
1029{
1030 int i;
1031 struct page *page;
1032
1033 i = (rbio->nr_data * rbio->stripe_len) >> PAGE_CACHE_SHIFT;
1034
1035 for (; i < rbio->nr_pages; i++) {
1036 if (rbio->stripe_pages[i])
1037 continue;
1038 page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1039 if (!page)
1040 return -ENOMEM;
1041 rbio->stripe_pages[i] = page;
1042 }
1043 return 0;
1044}
1045
1046/*
1047 * add a single page from a specific stripe into our list of bios for IO
1048 * this will try to merge into existing bios if possible, and returns
1049 * zero if all went well.
1050 */
Eric Sandeen48a3b632013-04-25 20:41:01 +00001051static int rbio_add_io_page(struct btrfs_raid_bio *rbio,
1052 struct bio_list *bio_list,
1053 struct page *page,
1054 int stripe_nr,
1055 unsigned long page_index,
1056 unsigned long bio_max_len)
David Woodhouse53b381b2013-01-29 18:40:14 -05001057{
1058 struct bio *last = bio_list->tail;
1059 u64 last_end = 0;
1060 int ret;
1061 struct bio *bio;
1062 struct btrfs_bio_stripe *stripe;
1063 u64 disk_start;
1064
1065 stripe = &rbio->bbio->stripes[stripe_nr];
1066 disk_start = stripe->physical + (page_index << PAGE_CACHE_SHIFT);
1067
1068 /* if the device is missing, just fail this stripe */
1069 if (!stripe->dev->bdev)
1070 return fail_rbio_index(rbio, stripe_nr);
1071
1072 /* see if we can add this page onto our existing bio */
1073 if (last) {
Kent Overstreet4f024f32013-10-11 15:44:27 -07001074 last_end = (u64)last->bi_iter.bi_sector << 9;
1075 last_end += last->bi_iter.bi_size;
David Woodhouse53b381b2013-01-29 18:40:14 -05001076
1077 /*
1078 * we can't merge these if they are from different
1079 * devices or if they are not contiguous
1080 */
1081 if (last_end == disk_start && stripe->dev->bdev &&
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001082 !last->bi_error &&
David Woodhouse53b381b2013-01-29 18:40:14 -05001083 last->bi_bdev == stripe->dev->bdev) {
1084 ret = bio_add_page(last, page, PAGE_CACHE_SIZE, 0);
1085 if (ret == PAGE_CACHE_SIZE)
1086 return 0;
1087 }
1088 }
1089
1090 /* put a new bio on the list */
Chris Mason9be33952013-05-17 18:30:14 -04001091 bio = btrfs_io_bio_alloc(GFP_NOFS, bio_max_len >> PAGE_SHIFT?:1);
David Woodhouse53b381b2013-01-29 18:40:14 -05001092 if (!bio)
1093 return -ENOMEM;
1094
Kent Overstreet4f024f32013-10-11 15:44:27 -07001095 bio->bi_iter.bi_size = 0;
David Woodhouse53b381b2013-01-29 18:40:14 -05001096 bio->bi_bdev = stripe->dev->bdev;
Kent Overstreet4f024f32013-10-11 15:44:27 -07001097 bio->bi_iter.bi_sector = disk_start >> 9;
David Woodhouse53b381b2013-01-29 18:40:14 -05001098
1099 bio_add_page(bio, page, PAGE_CACHE_SIZE, 0);
1100 bio_list_add(bio_list, bio);
1101 return 0;
1102}
1103
1104/*
1105 * while we're doing the read/modify/write cycle, we could
1106 * have errors in reading pages off the disk. This checks
1107 * for errors and if we're not able to read the page it'll
1108 * trigger parity reconstruction. The rmw will be finished
1109 * after we've reconstructed the failed stripes
1110 */
1111static void validate_rbio_for_rmw(struct btrfs_raid_bio *rbio)
1112{
1113 if (rbio->faila >= 0 || rbio->failb >= 0) {
Miao Xie2c8cdd62014-11-14 16:06:25 +08001114 BUG_ON(rbio->faila == rbio->real_stripes - 1);
David Woodhouse53b381b2013-01-29 18:40:14 -05001115 __raid56_parity_recover(rbio);
1116 } else {
1117 finish_rmw(rbio);
1118 }
1119}
1120
1121/*
1122 * these are just the pages from the rbio array, not from anything
1123 * the FS sent down to us
1124 */
1125static struct page *rbio_stripe_page(struct btrfs_raid_bio *rbio, int stripe, int page)
1126{
1127 int index;
1128 index = stripe * (rbio->stripe_len >> PAGE_CACHE_SHIFT);
1129 index += page;
1130 return rbio->stripe_pages[index];
1131}
1132
1133/*
1134 * helper function to walk our bio list and populate the bio_pages array with
1135 * the result. This seems expensive, but it is faster than constantly
1136 * searching through the bio list as we setup the IO in finish_rmw or stripe
1137 * reconstruction.
1138 *
1139 * This must be called before you trust the answers from page_in_rbio
1140 */
1141static void index_rbio_pages(struct btrfs_raid_bio *rbio)
1142{
1143 struct bio *bio;
1144 u64 start;
1145 unsigned long stripe_offset;
1146 unsigned long page_index;
1147 struct page *p;
1148 int i;
1149
1150 spin_lock_irq(&rbio->bio_list_lock);
1151 bio_list_for_each(bio, &rbio->bio_list) {
Kent Overstreet4f024f32013-10-11 15:44:27 -07001152 start = (u64)bio->bi_iter.bi_sector << 9;
Zhao Lei8e5cfb52015-01-20 15:11:33 +08001153 stripe_offset = start - rbio->bbio->raid_map[0];
David Woodhouse53b381b2013-01-29 18:40:14 -05001154 page_index = stripe_offset >> PAGE_CACHE_SHIFT;
1155
1156 for (i = 0; i < bio->bi_vcnt; i++) {
1157 p = bio->bi_io_vec[i].bv_page;
1158 rbio->bio_pages[page_index + i] = p;
1159 }
1160 }
1161 spin_unlock_irq(&rbio->bio_list_lock);
1162}
1163
1164/*
1165 * this is called from one of two situations. We either
1166 * have a full stripe from the higher layers, or we've read all
1167 * the missing bits off disk.
1168 *
1169 * This will calculate the parity and then send down any
1170 * changed blocks.
1171 */
1172static noinline void finish_rmw(struct btrfs_raid_bio *rbio)
1173{
1174 struct btrfs_bio *bbio = rbio->bbio;
Miao Xie2c8cdd62014-11-14 16:06:25 +08001175 void *pointers[rbio->real_stripes];
David Woodhouse53b381b2013-01-29 18:40:14 -05001176 int stripe_len = rbio->stripe_len;
1177 int nr_data = rbio->nr_data;
1178 int stripe;
1179 int pagenr;
1180 int p_stripe = -1;
1181 int q_stripe = -1;
1182 struct bio_list bio_list;
1183 struct bio *bio;
1184 int pages_per_stripe = stripe_len >> PAGE_CACHE_SHIFT;
1185 int ret;
1186
1187 bio_list_init(&bio_list);
1188
Miao Xie2c8cdd62014-11-14 16:06:25 +08001189 if (rbio->real_stripes - rbio->nr_data == 1) {
1190 p_stripe = rbio->real_stripes - 1;
1191 } else if (rbio->real_stripes - rbio->nr_data == 2) {
1192 p_stripe = rbio->real_stripes - 2;
1193 q_stripe = rbio->real_stripes - 1;
David Woodhouse53b381b2013-01-29 18:40:14 -05001194 } else {
1195 BUG();
1196 }
1197
1198 /* at this point we either have a full stripe,
1199 * or we've read the full stripe from the drive.
1200 * recalculate the parity and write the new results.
1201 *
1202 * We're not allowed to add any new bios to the
1203 * bio list here, anyone else that wants to
1204 * change this stripe needs to do their own rmw.
1205 */
1206 spin_lock_irq(&rbio->bio_list_lock);
1207 set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
1208 spin_unlock_irq(&rbio->bio_list_lock);
1209
Miao Xieb89e1b02014-10-15 11:18:44 +08001210 atomic_set(&rbio->error, 0);
David Woodhouse53b381b2013-01-29 18:40:14 -05001211
1212 /*
1213 * now that we've set rmw_locked, run through the
1214 * bio list one last time and map the page pointers
Chris Mason4ae10b32013-01-31 14:42:09 -05001215 *
1216 * We don't cache full rbios because we're assuming
1217 * the higher layers are unlikely to use this area of
1218 * the disk again soon. If they do use it again,
1219 * hopefully they will send another full bio.
David Woodhouse53b381b2013-01-29 18:40:14 -05001220 */
1221 index_rbio_pages(rbio);
Chris Mason4ae10b32013-01-31 14:42:09 -05001222 if (!rbio_is_full(rbio))
1223 cache_rbio_pages(rbio);
1224 else
1225 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
David Woodhouse53b381b2013-01-29 18:40:14 -05001226
1227 for (pagenr = 0; pagenr < pages_per_stripe; pagenr++) {
1228 struct page *p;
1229 /* first collect one page from each data stripe */
1230 for (stripe = 0; stripe < nr_data; stripe++) {
1231 p = page_in_rbio(rbio, stripe, pagenr, 0);
1232 pointers[stripe] = kmap(p);
1233 }
1234
1235 /* then add the parity stripe */
1236 p = rbio_pstripe_page(rbio, pagenr);
1237 SetPageUptodate(p);
1238 pointers[stripe++] = kmap(p);
1239
1240 if (q_stripe != -1) {
1241
1242 /*
1243 * raid6, add the qstripe and call the
1244 * library function to fill in our p/q
1245 */
1246 p = rbio_qstripe_page(rbio, pagenr);
1247 SetPageUptodate(p);
1248 pointers[stripe++] = kmap(p);
1249
Miao Xie2c8cdd62014-11-14 16:06:25 +08001250 raid6_call.gen_syndrome(rbio->real_stripes, PAGE_SIZE,
David Woodhouse53b381b2013-01-29 18:40:14 -05001251 pointers);
1252 } else {
1253 /* raid5 */
1254 memcpy(pointers[nr_data], pointers[0], PAGE_SIZE);
1255 run_xor(pointers + 1, nr_data - 1, PAGE_CACHE_SIZE);
1256 }
1257
1258
Miao Xie2c8cdd62014-11-14 16:06:25 +08001259 for (stripe = 0; stripe < rbio->real_stripes; stripe++)
David Woodhouse53b381b2013-01-29 18:40:14 -05001260 kunmap(page_in_rbio(rbio, stripe, pagenr, 0));
1261 }
1262
1263 /*
1264 * time to start writing. Make bios for everything from the
1265 * higher layers (the bio_list in our rbio) and our p/q. Ignore
1266 * everything else.
1267 */
Miao Xie2c8cdd62014-11-14 16:06:25 +08001268 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
David Woodhouse53b381b2013-01-29 18:40:14 -05001269 for (pagenr = 0; pagenr < pages_per_stripe; pagenr++) {
1270 struct page *page;
1271 if (stripe < rbio->nr_data) {
1272 page = page_in_rbio(rbio, stripe, pagenr, 1);
1273 if (!page)
1274 continue;
1275 } else {
1276 page = rbio_stripe_page(rbio, stripe, pagenr);
1277 }
1278
1279 ret = rbio_add_io_page(rbio, &bio_list,
1280 page, stripe, pagenr, rbio->stripe_len);
1281 if (ret)
1282 goto cleanup;
1283 }
1284 }
1285
Miao Xie2c8cdd62014-11-14 16:06:25 +08001286 if (likely(!bbio->num_tgtdevs))
1287 goto write_data;
1288
1289 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
1290 if (!bbio->tgtdev_map[stripe])
1291 continue;
1292
1293 for (pagenr = 0; pagenr < pages_per_stripe; pagenr++) {
1294 struct page *page;
1295 if (stripe < rbio->nr_data) {
1296 page = page_in_rbio(rbio, stripe, pagenr, 1);
1297 if (!page)
1298 continue;
1299 } else {
1300 page = rbio_stripe_page(rbio, stripe, pagenr);
1301 }
1302
1303 ret = rbio_add_io_page(rbio, &bio_list, page,
1304 rbio->bbio->tgtdev_map[stripe],
1305 pagenr, rbio->stripe_len);
1306 if (ret)
1307 goto cleanup;
1308 }
1309 }
1310
1311write_data:
Miao Xieb89e1b02014-10-15 11:18:44 +08001312 atomic_set(&rbio->stripes_pending, bio_list_size(&bio_list));
1313 BUG_ON(atomic_read(&rbio->stripes_pending) == 0);
David Woodhouse53b381b2013-01-29 18:40:14 -05001314
1315 while (1) {
1316 bio = bio_list_pop(&bio_list);
1317 if (!bio)
1318 break;
1319
1320 bio->bi_private = rbio;
1321 bio->bi_end_io = raid_write_end_io;
David Woodhouse53b381b2013-01-29 18:40:14 -05001322 submit_bio(WRITE, bio);
1323 }
1324 return;
1325
1326cleanup:
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001327 rbio_orig_end_io(rbio, -EIO);
David Woodhouse53b381b2013-01-29 18:40:14 -05001328}
1329
1330/*
1331 * helper to find the stripe number for a given bio. Used to figure out which
1332 * stripe has failed. This expects the bio to correspond to a physical disk,
1333 * so it looks up based on physical sector numbers.
1334 */
1335static int find_bio_stripe(struct btrfs_raid_bio *rbio,
1336 struct bio *bio)
1337{
Kent Overstreet4f024f32013-10-11 15:44:27 -07001338 u64 physical = bio->bi_iter.bi_sector;
David Woodhouse53b381b2013-01-29 18:40:14 -05001339 u64 stripe_start;
1340 int i;
1341 struct btrfs_bio_stripe *stripe;
1342
1343 physical <<= 9;
1344
1345 for (i = 0; i < rbio->bbio->num_stripes; i++) {
1346 stripe = &rbio->bbio->stripes[i];
1347 stripe_start = stripe->physical;
1348 if (physical >= stripe_start &&
Miao Xie2c8cdd62014-11-14 16:06:25 +08001349 physical < stripe_start + rbio->stripe_len &&
1350 bio->bi_bdev == stripe->dev->bdev) {
David Woodhouse53b381b2013-01-29 18:40:14 -05001351 return i;
1352 }
1353 }
1354 return -1;
1355}
1356
1357/*
1358 * helper to find the stripe number for a given
1359 * bio (before mapping). Used to figure out which stripe has
1360 * failed. This looks up based on logical block numbers.
1361 */
1362static int find_logical_bio_stripe(struct btrfs_raid_bio *rbio,
1363 struct bio *bio)
1364{
Kent Overstreet4f024f32013-10-11 15:44:27 -07001365 u64 logical = bio->bi_iter.bi_sector;
David Woodhouse53b381b2013-01-29 18:40:14 -05001366 u64 stripe_start;
1367 int i;
1368
1369 logical <<= 9;
1370
1371 for (i = 0; i < rbio->nr_data; i++) {
Zhao Lei8e5cfb52015-01-20 15:11:33 +08001372 stripe_start = rbio->bbio->raid_map[i];
David Woodhouse53b381b2013-01-29 18:40:14 -05001373 if (logical >= stripe_start &&
1374 logical < stripe_start + rbio->stripe_len) {
1375 return i;
1376 }
1377 }
1378 return -1;
1379}
1380
1381/*
1382 * returns -EIO if we had too many failures
1383 */
1384static int fail_rbio_index(struct btrfs_raid_bio *rbio, int failed)
1385{
1386 unsigned long flags;
1387 int ret = 0;
1388
1389 spin_lock_irqsave(&rbio->bio_list_lock, flags);
1390
1391 /* we already know this stripe is bad, move on */
1392 if (rbio->faila == failed || rbio->failb == failed)
1393 goto out;
1394
1395 if (rbio->faila == -1) {
1396 /* first failure on this rbio */
1397 rbio->faila = failed;
Miao Xieb89e1b02014-10-15 11:18:44 +08001398 atomic_inc(&rbio->error);
David Woodhouse53b381b2013-01-29 18:40:14 -05001399 } else if (rbio->failb == -1) {
1400 /* second failure on this rbio */
1401 rbio->failb = failed;
Miao Xieb89e1b02014-10-15 11:18:44 +08001402 atomic_inc(&rbio->error);
David Woodhouse53b381b2013-01-29 18:40:14 -05001403 } else {
1404 ret = -EIO;
1405 }
1406out:
1407 spin_unlock_irqrestore(&rbio->bio_list_lock, flags);
1408
1409 return ret;
1410}
1411
1412/*
1413 * helper to fail a stripe based on a physical disk
1414 * bio.
1415 */
1416static int fail_bio_stripe(struct btrfs_raid_bio *rbio,
1417 struct bio *bio)
1418{
1419 int failed = find_bio_stripe(rbio, bio);
1420
1421 if (failed < 0)
1422 return -EIO;
1423
1424 return fail_rbio_index(rbio, failed);
1425}
1426
1427/*
1428 * this sets each page in the bio uptodate. It should only be used on private
1429 * rbio pages, nothing that comes in from the higher layers
1430 */
1431static void set_bio_pages_uptodate(struct bio *bio)
1432{
1433 int i;
1434 struct page *p;
1435
1436 for (i = 0; i < bio->bi_vcnt; i++) {
1437 p = bio->bi_io_vec[i].bv_page;
1438 SetPageUptodate(p);
1439 }
1440}
1441
1442/*
1443 * end io for the read phase of the rmw cycle. All the bios here are physical
1444 * stripe bios we've read from the disk so we can recalculate the parity of the
1445 * stripe.
1446 *
1447 * This will usually kick off finish_rmw once all the bios are read in, but it
1448 * may trigger parity reconstruction if we had any errors along the way
1449 */
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001450static void raid_rmw_end_io(struct bio *bio)
David Woodhouse53b381b2013-01-29 18:40:14 -05001451{
1452 struct btrfs_raid_bio *rbio = bio->bi_private;
1453
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001454 if (bio->bi_error)
David Woodhouse53b381b2013-01-29 18:40:14 -05001455 fail_bio_stripe(rbio, bio);
1456 else
1457 set_bio_pages_uptodate(bio);
1458
1459 bio_put(bio);
1460
Miao Xieb89e1b02014-10-15 11:18:44 +08001461 if (!atomic_dec_and_test(&rbio->stripes_pending))
David Woodhouse53b381b2013-01-29 18:40:14 -05001462 return;
1463
Miao Xieb89e1b02014-10-15 11:18:44 +08001464 if (atomic_read(&rbio->error) > rbio->bbio->max_errors)
David Woodhouse53b381b2013-01-29 18:40:14 -05001465 goto cleanup;
1466
1467 /*
1468 * this will normally call finish_rmw to start our write
1469 * but if there are any failed stripes we'll reconstruct
1470 * from parity first
1471 */
1472 validate_rbio_for_rmw(rbio);
1473 return;
1474
1475cleanup:
1476
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001477 rbio_orig_end_io(rbio, -EIO);
David Woodhouse53b381b2013-01-29 18:40:14 -05001478}
1479
1480static void async_rmw_stripe(struct btrfs_raid_bio *rbio)
1481{
Liu Bo9e0af232014-08-15 23:36:53 +08001482 btrfs_init_work(&rbio->work, btrfs_rmw_helper,
1483 rmw_work, NULL, NULL);
David Woodhouse53b381b2013-01-29 18:40:14 -05001484
Qu Wenruod05a33a2014-02-28 10:46:11 +08001485 btrfs_queue_work(rbio->fs_info->rmw_workers,
1486 &rbio->work);
David Woodhouse53b381b2013-01-29 18:40:14 -05001487}
1488
1489static void async_read_rebuild(struct btrfs_raid_bio *rbio)
1490{
Liu Bo9e0af232014-08-15 23:36:53 +08001491 btrfs_init_work(&rbio->work, btrfs_rmw_helper,
1492 read_rebuild_work, NULL, NULL);
David Woodhouse53b381b2013-01-29 18:40:14 -05001493
Qu Wenruod05a33a2014-02-28 10:46:11 +08001494 btrfs_queue_work(rbio->fs_info->rmw_workers,
1495 &rbio->work);
David Woodhouse53b381b2013-01-29 18:40:14 -05001496}
1497
1498/*
1499 * the stripe must be locked by the caller. It will
1500 * unlock after all the writes are done
1501 */
1502static int raid56_rmw_stripe(struct btrfs_raid_bio *rbio)
1503{
1504 int bios_to_read = 0;
David Woodhouse53b381b2013-01-29 18:40:14 -05001505 struct bio_list bio_list;
1506 int ret;
David Sterbaed6078f2014-06-05 01:59:57 +02001507 int nr_pages = DIV_ROUND_UP(rbio->stripe_len, PAGE_CACHE_SIZE);
David Woodhouse53b381b2013-01-29 18:40:14 -05001508 int pagenr;
1509 int stripe;
1510 struct bio *bio;
1511
1512 bio_list_init(&bio_list);
1513
1514 ret = alloc_rbio_pages(rbio);
1515 if (ret)
1516 goto cleanup;
1517
1518 index_rbio_pages(rbio);
1519
Miao Xieb89e1b02014-10-15 11:18:44 +08001520 atomic_set(&rbio->error, 0);
David Woodhouse53b381b2013-01-29 18:40:14 -05001521 /*
1522 * build a list of bios to read all the missing parts of this
1523 * stripe
1524 */
1525 for (stripe = 0; stripe < rbio->nr_data; stripe++) {
1526 for (pagenr = 0; pagenr < nr_pages; pagenr++) {
1527 struct page *page;
1528 /*
1529 * we want to find all the pages missing from
1530 * the rbio and read them from the disk. If
1531 * page_in_rbio finds a page in the bio list
1532 * we don't need to read it off the stripe.
1533 */
1534 page = page_in_rbio(rbio, stripe, pagenr, 1);
1535 if (page)
1536 continue;
1537
1538 page = rbio_stripe_page(rbio, stripe, pagenr);
Chris Mason4ae10b32013-01-31 14:42:09 -05001539 /*
1540 * the bio cache may have handed us an uptodate
1541 * page. If so, be happy and use it
1542 */
1543 if (PageUptodate(page))
1544 continue;
1545
David Woodhouse53b381b2013-01-29 18:40:14 -05001546 ret = rbio_add_io_page(rbio, &bio_list, page,
1547 stripe, pagenr, rbio->stripe_len);
1548 if (ret)
1549 goto cleanup;
1550 }
1551 }
1552
1553 bios_to_read = bio_list_size(&bio_list);
1554 if (!bios_to_read) {
1555 /*
1556 * this can happen if others have merged with
1557 * us, it means there is nothing left to read.
1558 * But if there are missing devices it may not be
1559 * safe to do the full stripe write yet.
1560 */
1561 goto finish;
1562 }
1563
1564 /*
1565 * the bbio may be freed once we submit the last bio. Make sure
1566 * not to touch it after that
1567 */
Miao Xieb89e1b02014-10-15 11:18:44 +08001568 atomic_set(&rbio->stripes_pending, bios_to_read);
David Woodhouse53b381b2013-01-29 18:40:14 -05001569 while (1) {
1570 bio = bio_list_pop(&bio_list);
1571 if (!bio)
1572 break;
1573
1574 bio->bi_private = rbio;
1575 bio->bi_end_io = raid_rmw_end_io;
1576
1577 btrfs_bio_wq_end_io(rbio->fs_info, bio,
1578 BTRFS_WQ_ENDIO_RAID56);
1579
David Woodhouse53b381b2013-01-29 18:40:14 -05001580 submit_bio(READ, bio);
1581 }
1582 /* the actual write will happen once the reads are done */
1583 return 0;
1584
1585cleanup:
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001586 rbio_orig_end_io(rbio, -EIO);
David Woodhouse53b381b2013-01-29 18:40:14 -05001587 return -EIO;
1588
1589finish:
1590 validate_rbio_for_rmw(rbio);
1591 return 0;
1592}
1593
1594/*
1595 * if the upper layers pass in a full stripe, we thank them by only allocating
1596 * enough pages to hold the parity, and sending it all down quickly.
1597 */
1598static int full_stripe_write(struct btrfs_raid_bio *rbio)
1599{
1600 int ret;
1601
1602 ret = alloc_rbio_parity_pages(rbio);
Miao Xie3cd846d2013-07-22 16:36:57 +08001603 if (ret) {
1604 __free_raid_bio(rbio);
David Woodhouse53b381b2013-01-29 18:40:14 -05001605 return ret;
Miao Xie3cd846d2013-07-22 16:36:57 +08001606 }
David Woodhouse53b381b2013-01-29 18:40:14 -05001607
1608 ret = lock_stripe_add(rbio);
1609 if (ret == 0)
1610 finish_rmw(rbio);
1611 return 0;
1612}
1613
1614/*
1615 * partial stripe writes get handed over to async helpers.
1616 * We're really hoping to merge a few more writes into this
1617 * rbio before calculating new parity
1618 */
1619static int partial_stripe_write(struct btrfs_raid_bio *rbio)
1620{
1621 int ret;
1622
1623 ret = lock_stripe_add(rbio);
1624 if (ret == 0)
1625 async_rmw_stripe(rbio);
1626 return 0;
1627}
1628
1629/*
1630 * sometimes while we were reading from the drive to
1631 * recalculate parity, enough new bios come into create
1632 * a full stripe. So we do a check here to see if we can
1633 * go directly to finish_rmw
1634 */
1635static int __raid56_parity_write(struct btrfs_raid_bio *rbio)
1636{
1637 /* head off into rmw land if we don't have a full stripe */
1638 if (!rbio_is_full(rbio))
1639 return partial_stripe_write(rbio);
1640 return full_stripe_write(rbio);
1641}
1642
1643/*
Chris Mason6ac0f482013-01-31 14:42:28 -05001644 * We use plugging call backs to collect full stripes.
1645 * Any time we get a partial stripe write while plugged
1646 * we collect it into a list. When the unplug comes down,
1647 * we sort the list by logical block number and merge
1648 * everything we can into the same rbios
1649 */
1650struct btrfs_plug_cb {
1651 struct blk_plug_cb cb;
1652 struct btrfs_fs_info *info;
1653 struct list_head rbio_list;
1654 struct btrfs_work work;
1655};
1656
1657/*
1658 * rbios on the plug list are sorted for easier merging.
1659 */
1660static int plug_cmp(void *priv, struct list_head *a, struct list_head *b)
1661{
1662 struct btrfs_raid_bio *ra = container_of(a, struct btrfs_raid_bio,
1663 plug_list);
1664 struct btrfs_raid_bio *rb = container_of(b, struct btrfs_raid_bio,
1665 plug_list);
Kent Overstreet4f024f32013-10-11 15:44:27 -07001666 u64 a_sector = ra->bio_list.head->bi_iter.bi_sector;
1667 u64 b_sector = rb->bio_list.head->bi_iter.bi_sector;
Chris Mason6ac0f482013-01-31 14:42:28 -05001668
1669 if (a_sector < b_sector)
1670 return -1;
1671 if (a_sector > b_sector)
1672 return 1;
1673 return 0;
1674}
1675
1676static void run_plug(struct btrfs_plug_cb *plug)
1677{
1678 struct btrfs_raid_bio *cur;
1679 struct btrfs_raid_bio *last = NULL;
1680
1681 /*
1682 * sort our plug list then try to merge
1683 * everything we can in hopes of creating full
1684 * stripes.
1685 */
1686 list_sort(NULL, &plug->rbio_list, plug_cmp);
1687 while (!list_empty(&plug->rbio_list)) {
1688 cur = list_entry(plug->rbio_list.next,
1689 struct btrfs_raid_bio, plug_list);
1690 list_del_init(&cur->plug_list);
1691
1692 if (rbio_is_full(cur)) {
1693 /* we have a full stripe, send it down */
1694 full_stripe_write(cur);
1695 continue;
1696 }
1697 if (last) {
1698 if (rbio_can_merge(last, cur)) {
1699 merge_rbio(last, cur);
1700 __free_raid_bio(cur);
1701 continue;
1702
1703 }
1704 __raid56_parity_write(last);
1705 }
1706 last = cur;
1707 }
1708 if (last) {
1709 __raid56_parity_write(last);
1710 }
1711 kfree(plug);
1712}
1713
1714/*
1715 * if the unplug comes from schedule, we have to push the
1716 * work off to a helper thread
1717 */
1718static void unplug_work(struct btrfs_work *work)
1719{
1720 struct btrfs_plug_cb *plug;
1721 plug = container_of(work, struct btrfs_plug_cb, work);
1722 run_plug(plug);
1723}
1724
1725static void btrfs_raid_unplug(struct blk_plug_cb *cb, bool from_schedule)
1726{
1727 struct btrfs_plug_cb *plug;
1728 plug = container_of(cb, struct btrfs_plug_cb, cb);
1729
1730 if (from_schedule) {
Liu Bo9e0af232014-08-15 23:36:53 +08001731 btrfs_init_work(&plug->work, btrfs_rmw_helper,
1732 unplug_work, NULL, NULL);
Qu Wenruod05a33a2014-02-28 10:46:11 +08001733 btrfs_queue_work(plug->info->rmw_workers,
1734 &plug->work);
Chris Mason6ac0f482013-01-31 14:42:28 -05001735 return;
1736 }
1737 run_plug(plug);
1738}
1739
1740/*
David Woodhouse53b381b2013-01-29 18:40:14 -05001741 * our main entry point for writes from the rest of the FS.
1742 */
1743int raid56_parity_write(struct btrfs_root *root, struct bio *bio,
Zhao Lei8e5cfb52015-01-20 15:11:33 +08001744 struct btrfs_bio *bbio, u64 stripe_len)
David Woodhouse53b381b2013-01-29 18:40:14 -05001745{
1746 struct btrfs_raid_bio *rbio;
Chris Mason6ac0f482013-01-31 14:42:28 -05001747 struct btrfs_plug_cb *plug = NULL;
1748 struct blk_plug_cb *cb;
Miao Xie42452152014-11-25 16:39:28 +08001749 int ret;
David Woodhouse53b381b2013-01-29 18:40:14 -05001750
Zhao Lei8e5cfb52015-01-20 15:11:33 +08001751 rbio = alloc_rbio(root, bbio, stripe_len);
Miao Xieaf8e2d12014-10-23 14:42:50 +08001752 if (IS_ERR(rbio)) {
Zhao Lei6e9606d2015-01-20 15:11:34 +08001753 btrfs_put_bbio(bbio);
David Woodhouse53b381b2013-01-29 18:40:14 -05001754 return PTR_ERR(rbio);
Miao Xieaf8e2d12014-10-23 14:42:50 +08001755 }
David Woodhouse53b381b2013-01-29 18:40:14 -05001756 bio_list_add(&rbio->bio_list, bio);
Kent Overstreet4f024f32013-10-11 15:44:27 -07001757 rbio->bio_list_bytes = bio->bi_iter.bi_size;
Miao Xie1b94b552014-11-06 16:14:21 +08001758 rbio->operation = BTRFS_RBIO_WRITE;
Chris Mason6ac0f482013-01-31 14:42:28 -05001759
Miao Xie42452152014-11-25 16:39:28 +08001760 btrfs_bio_counter_inc_noblocked(root->fs_info);
1761 rbio->generic_bio_cnt = 1;
1762
Chris Mason6ac0f482013-01-31 14:42:28 -05001763 /*
1764 * don't plug on full rbios, just get them out the door
1765 * as quickly as we can
1766 */
Miao Xie42452152014-11-25 16:39:28 +08001767 if (rbio_is_full(rbio)) {
1768 ret = full_stripe_write(rbio);
1769 if (ret)
1770 btrfs_bio_counter_dec(root->fs_info);
1771 return ret;
1772 }
Chris Mason6ac0f482013-01-31 14:42:28 -05001773
1774 cb = blk_check_plugged(btrfs_raid_unplug, root->fs_info,
1775 sizeof(*plug));
1776 if (cb) {
1777 plug = container_of(cb, struct btrfs_plug_cb, cb);
1778 if (!plug->info) {
1779 plug->info = root->fs_info;
1780 INIT_LIST_HEAD(&plug->rbio_list);
1781 }
1782 list_add_tail(&rbio->plug_list, &plug->rbio_list);
Miao Xie42452152014-11-25 16:39:28 +08001783 ret = 0;
Chris Mason6ac0f482013-01-31 14:42:28 -05001784 } else {
Miao Xie42452152014-11-25 16:39:28 +08001785 ret = __raid56_parity_write(rbio);
1786 if (ret)
1787 btrfs_bio_counter_dec(root->fs_info);
Chris Mason6ac0f482013-01-31 14:42:28 -05001788 }
Miao Xie42452152014-11-25 16:39:28 +08001789 return ret;
David Woodhouse53b381b2013-01-29 18:40:14 -05001790}
1791
1792/*
1793 * all parity reconstruction happens here. We've read in everything
1794 * we can find from the drives and this does the heavy lifting of
1795 * sorting the good from the bad.
1796 */
1797static void __raid_recover_end_io(struct btrfs_raid_bio *rbio)
1798{
1799 int pagenr, stripe;
1800 void **pointers;
1801 int faila = -1, failb = -1;
David Sterbaed6078f2014-06-05 01:59:57 +02001802 int nr_pages = DIV_ROUND_UP(rbio->stripe_len, PAGE_CACHE_SIZE);
David Woodhouse53b381b2013-01-29 18:40:14 -05001803 struct page *page;
1804 int err;
1805 int i;
1806
David Sterba31e818f2015-02-20 18:00:26 +01001807 pointers = kcalloc(rbio->real_stripes, sizeof(void *), GFP_NOFS);
David Woodhouse53b381b2013-01-29 18:40:14 -05001808 if (!pointers) {
1809 err = -ENOMEM;
1810 goto cleanup_io;
1811 }
1812
1813 faila = rbio->faila;
1814 failb = rbio->failb;
1815
Omar Sandovalb4ee1782015-06-19 11:52:50 -07001816 if (rbio->operation == BTRFS_RBIO_READ_REBUILD ||
1817 rbio->operation == BTRFS_RBIO_REBUILD_MISSING) {
David Woodhouse53b381b2013-01-29 18:40:14 -05001818 spin_lock_irq(&rbio->bio_list_lock);
1819 set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
1820 spin_unlock_irq(&rbio->bio_list_lock);
1821 }
1822
1823 index_rbio_pages(rbio);
1824
1825 for (pagenr = 0; pagenr < nr_pages; pagenr++) {
Miao Xie5a6ac9e2014-11-06 17:20:58 +08001826 /*
1827 * Now we just use bitmap to mark the horizontal stripes in
1828 * which we have data when doing parity scrub.
1829 */
1830 if (rbio->operation == BTRFS_RBIO_PARITY_SCRUB &&
1831 !test_bit(pagenr, rbio->dbitmap))
1832 continue;
1833
David Woodhouse53b381b2013-01-29 18:40:14 -05001834 /* setup our array of pointers with pages
1835 * from each stripe
1836 */
Miao Xie2c8cdd62014-11-14 16:06:25 +08001837 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
David Woodhouse53b381b2013-01-29 18:40:14 -05001838 /*
1839 * if we're rebuilding a read, we have to use
1840 * pages from the bio list
1841 */
Omar Sandovalb4ee1782015-06-19 11:52:50 -07001842 if ((rbio->operation == BTRFS_RBIO_READ_REBUILD ||
1843 rbio->operation == BTRFS_RBIO_REBUILD_MISSING) &&
David Woodhouse53b381b2013-01-29 18:40:14 -05001844 (stripe == faila || stripe == failb)) {
1845 page = page_in_rbio(rbio, stripe, pagenr, 0);
1846 } else {
1847 page = rbio_stripe_page(rbio, stripe, pagenr);
1848 }
1849 pointers[stripe] = kmap(page);
1850 }
1851
1852 /* all raid6 handling here */
Zhao Lei10f11902015-01-20 15:11:43 +08001853 if (rbio->bbio->map_type & BTRFS_BLOCK_GROUP_RAID6) {
David Woodhouse53b381b2013-01-29 18:40:14 -05001854 /*
1855 * single failure, rebuild from parity raid5
1856 * style
1857 */
1858 if (failb < 0) {
1859 if (faila == rbio->nr_data) {
1860 /*
1861 * Just the P stripe has failed, without
1862 * a bad data or Q stripe.
1863 * TODO, we should redo the xor here.
1864 */
1865 err = -EIO;
1866 goto cleanup;
1867 }
1868 /*
1869 * a single failure in raid6 is rebuilt
1870 * in the pstripe code below
1871 */
1872 goto pstripe;
1873 }
1874
1875 /* make sure our ps and qs are in order */
1876 if (faila > failb) {
1877 int tmp = failb;
1878 failb = faila;
1879 faila = tmp;
1880 }
1881
1882 /* if the q stripe is failed, do a pstripe reconstruction
1883 * from the xors.
1884 * If both the q stripe and the P stripe are failed, we're
1885 * here due to a crc mismatch and we can't give them the
1886 * data they want
1887 */
Zhao Lei8e5cfb52015-01-20 15:11:33 +08001888 if (rbio->bbio->raid_map[failb] == RAID6_Q_STRIPE) {
1889 if (rbio->bbio->raid_map[faila] ==
1890 RAID5_P_STRIPE) {
David Woodhouse53b381b2013-01-29 18:40:14 -05001891 err = -EIO;
1892 goto cleanup;
1893 }
1894 /*
1895 * otherwise we have one bad data stripe and
1896 * a good P stripe. raid5!
1897 */
1898 goto pstripe;
1899 }
1900
Zhao Lei8e5cfb52015-01-20 15:11:33 +08001901 if (rbio->bbio->raid_map[failb] == RAID5_P_STRIPE) {
Miao Xie2c8cdd62014-11-14 16:06:25 +08001902 raid6_datap_recov(rbio->real_stripes,
David Woodhouse53b381b2013-01-29 18:40:14 -05001903 PAGE_SIZE, faila, pointers);
1904 } else {
Miao Xie2c8cdd62014-11-14 16:06:25 +08001905 raid6_2data_recov(rbio->real_stripes,
David Woodhouse53b381b2013-01-29 18:40:14 -05001906 PAGE_SIZE, faila, failb,
1907 pointers);
1908 }
1909 } else {
1910 void *p;
1911
1912 /* rebuild from P stripe here (raid5 or raid6) */
1913 BUG_ON(failb != -1);
1914pstripe:
1915 /* Copy parity block into failed block to start with */
1916 memcpy(pointers[faila],
1917 pointers[rbio->nr_data],
1918 PAGE_CACHE_SIZE);
1919
1920 /* rearrange the pointer array */
1921 p = pointers[faila];
1922 for (stripe = faila; stripe < rbio->nr_data - 1; stripe++)
1923 pointers[stripe] = pointers[stripe + 1];
1924 pointers[rbio->nr_data - 1] = p;
1925
1926 /* xor in the rest */
1927 run_xor(pointers, rbio->nr_data - 1, PAGE_CACHE_SIZE);
1928 }
1929 /* if we're doing this rebuild as part of an rmw, go through
1930 * and set all of our private rbio pages in the
1931 * failed stripes as uptodate. This way finish_rmw will
1932 * know they can be trusted. If this was a read reconstruction,
1933 * other endio functions will fiddle the uptodate bits
1934 */
Miao Xie1b94b552014-11-06 16:14:21 +08001935 if (rbio->operation == BTRFS_RBIO_WRITE) {
David Woodhouse53b381b2013-01-29 18:40:14 -05001936 for (i = 0; i < nr_pages; i++) {
1937 if (faila != -1) {
1938 page = rbio_stripe_page(rbio, faila, i);
1939 SetPageUptodate(page);
1940 }
1941 if (failb != -1) {
1942 page = rbio_stripe_page(rbio, failb, i);
1943 SetPageUptodate(page);
1944 }
1945 }
1946 }
Miao Xie2c8cdd62014-11-14 16:06:25 +08001947 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
David Woodhouse53b381b2013-01-29 18:40:14 -05001948 /*
1949 * if we're rebuilding a read, we have to use
1950 * pages from the bio list
1951 */
Omar Sandovalb4ee1782015-06-19 11:52:50 -07001952 if ((rbio->operation == BTRFS_RBIO_READ_REBUILD ||
1953 rbio->operation == BTRFS_RBIO_REBUILD_MISSING) &&
David Woodhouse53b381b2013-01-29 18:40:14 -05001954 (stripe == faila || stripe == failb)) {
1955 page = page_in_rbio(rbio, stripe, pagenr, 0);
1956 } else {
1957 page = rbio_stripe_page(rbio, stripe, pagenr);
1958 }
1959 kunmap(page);
1960 }
1961 }
1962
1963 err = 0;
1964cleanup:
1965 kfree(pointers);
1966
1967cleanup_io:
Miao Xie1b94b552014-11-06 16:14:21 +08001968 if (rbio->operation == BTRFS_RBIO_READ_REBUILD) {
Zhao Lei6e9606d2015-01-20 15:11:34 +08001969 if (err == 0)
Chris Mason4ae10b32013-01-31 14:42:09 -05001970 cache_rbio_pages(rbio);
1971 else
1972 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
1973
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001974 rbio_orig_end_io(rbio, err);
Omar Sandovalb4ee1782015-06-19 11:52:50 -07001975 } else if (rbio->operation == BTRFS_RBIO_REBUILD_MISSING) {
Linus Torvalds22365972015-09-05 15:14:43 -07001976 rbio_orig_end_io(rbio, err);
David Woodhouse53b381b2013-01-29 18:40:14 -05001977 } else if (err == 0) {
1978 rbio->faila = -1;
1979 rbio->failb = -1;
Miao Xie5a6ac9e2014-11-06 17:20:58 +08001980
1981 if (rbio->operation == BTRFS_RBIO_WRITE)
1982 finish_rmw(rbio);
1983 else if (rbio->operation == BTRFS_RBIO_PARITY_SCRUB)
1984 finish_parity_scrub(rbio, 0);
1985 else
1986 BUG();
David Woodhouse53b381b2013-01-29 18:40:14 -05001987 } else {
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001988 rbio_orig_end_io(rbio, err);
David Woodhouse53b381b2013-01-29 18:40:14 -05001989 }
1990}
1991
1992/*
1993 * This is called only for stripes we've read from disk to
1994 * reconstruct the parity.
1995 */
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001996static void raid_recover_end_io(struct bio *bio)
David Woodhouse53b381b2013-01-29 18:40:14 -05001997{
1998 struct btrfs_raid_bio *rbio = bio->bi_private;
1999
2000 /*
2001 * we only read stripe pages off the disk, set them
2002 * up to date if there were no errors
2003 */
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02002004 if (bio->bi_error)
David Woodhouse53b381b2013-01-29 18:40:14 -05002005 fail_bio_stripe(rbio, bio);
2006 else
2007 set_bio_pages_uptodate(bio);
2008 bio_put(bio);
2009
Miao Xieb89e1b02014-10-15 11:18:44 +08002010 if (!atomic_dec_and_test(&rbio->stripes_pending))
David Woodhouse53b381b2013-01-29 18:40:14 -05002011 return;
2012
Miao Xieb89e1b02014-10-15 11:18:44 +08002013 if (atomic_read(&rbio->error) > rbio->bbio->max_errors)
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02002014 rbio_orig_end_io(rbio, -EIO);
David Woodhouse53b381b2013-01-29 18:40:14 -05002015 else
2016 __raid_recover_end_io(rbio);
2017}
2018
2019/*
2020 * reads everything we need off the disk to reconstruct
2021 * the parity. endio handlers trigger final reconstruction
2022 * when the IO is done.
2023 *
2024 * This is used both for reads from the higher layers and for
2025 * parity construction required to finish a rmw cycle.
2026 */
2027static int __raid56_parity_recover(struct btrfs_raid_bio *rbio)
2028{
2029 int bios_to_read = 0;
David Woodhouse53b381b2013-01-29 18:40:14 -05002030 struct bio_list bio_list;
2031 int ret;
David Sterbaed6078f2014-06-05 01:59:57 +02002032 int nr_pages = DIV_ROUND_UP(rbio->stripe_len, PAGE_CACHE_SIZE);
David Woodhouse53b381b2013-01-29 18:40:14 -05002033 int pagenr;
2034 int stripe;
2035 struct bio *bio;
2036
2037 bio_list_init(&bio_list);
2038
2039 ret = alloc_rbio_pages(rbio);
2040 if (ret)
2041 goto cleanup;
2042
Miao Xieb89e1b02014-10-15 11:18:44 +08002043 atomic_set(&rbio->error, 0);
David Woodhouse53b381b2013-01-29 18:40:14 -05002044
2045 /*
Chris Mason4ae10b32013-01-31 14:42:09 -05002046 * read everything that hasn't failed. Thanks to the
2047 * stripe cache, it is possible that some or all of these
2048 * pages are going to be uptodate.
David Woodhouse53b381b2013-01-29 18:40:14 -05002049 */
Miao Xie2c8cdd62014-11-14 16:06:25 +08002050 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
Liu Bo55883832014-06-24 15:39:16 +08002051 if (rbio->faila == stripe || rbio->failb == stripe) {
Miao Xieb89e1b02014-10-15 11:18:44 +08002052 atomic_inc(&rbio->error);
David Woodhouse53b381b2013-01-29 18:40:14 -05002053 continue;
Liu Bo55883832014-06-24 15:39:16 +08002054 }
David Woodhouse53b381b2013-01-29 18:40:14 -05002055
2056 for (pagenr = 0; pagenr < nr_pages; pagenr++) {
2057 struct page *p;
2058
2059 /*
2060 * the rmw code may have already read this
2061 * page in
2062 */
2063 p = rbio_stripe_page(rbio, stripe, pagenr);
2064 if (PageUptodate(p))
2065 continue;
2066
2067 ret = rbio_add_io_page(rbio, &bio_list,
2068 rbio_stripe_page(rbio, stripe, pagenr),
2069 stripe, pagenr, rbio->stripe_len);
2070 if (ret < 0)
2071 goto cleanup;
2072 }
2073 }
2074
2075 bios_to_read = bio_list_size(&bio_list);
2076 if (!bios_to_read) {
2077 /*
2078 * we might have no bios to read just because the pages
2079 * were up to date, or we might have no bios to read because
2080 * the devices were gone.
2081 */
Miao Xieb89e1b02014-10-15 11:18:44 +08002082 if (atomic_read(&rbio->error) <= rbio->bbio->max_errors) {
David Woodhouse53b381b2013-01-29 18:40:14 -05002083 __raid_recover_end_io(rbio);
2084 goto out;
2085 } else {
2086 goto cleanup;
2087 }
2088 }
2089
2090 /*
2091 * the bbio may be freed once we submit the last bio. Make sure
2092 * not to touch it after that
2093 */
Miao Xieb89e1b02014-10-15 11:18:44 +08002094 atomic_set(&rbio->stripes_pending, bios_to_read);
David Woodhouse53b381b2013-01-29 18:40:14 -05002095 while (1) {
2096 bio = bio_list_pop(&bio_list);
2097 if (!bio)
2098 break;
2099
2100 bio->bi_private = rbio;
2101 bio->bi_end_io = raid_recover_end_io;
2102
2103 btrfs_bio_wq_end_io(rbio->fs_info, bio,
2104 BTRFS_WQ_ENDIO_RAID56);
2105
David Woodhouse53b381b2013-01-29 18:40:14 -05002106 submit_bio(READ, bio);
2107 }
2108out:
2109 return 0;
2110
2111cleanup:
Omar Sandovalb4ee1782015-06-19 11:52:50 -07002112 if (rbio->operation == BTRFS_RBIO_READ_REBUILD ||
2113 rbio->operation == BTRFS_RBIO_REBUILD_MISSING)
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02002114 rbio_orig_end_io(rbio, -EIO);
David Woodhouse53b381b2013-01-29 18:40:14 -05002115 return -EIO;
2116}
2117
2118/*
2119 * the main entry point for reads from the higher layers. This
2120 * is really only called when the normal read path had a failure,
2121 * so we assume the bio they send down corresponds to a failed part
2122 * of the drive.
2123 */
2124int raid56_parity_recover(struct btrfs_root *root, struct bio *bio,
Zhao Lei8e5cfb52015-01-20 15:11:33 +08002125 struct btrfs_bio *bbio, u64 stripe_len,
2126 int mirror_num, int generic_io)
David Woodhouse53b381b2013-01-29 18:40:14 -05002127{
2128 struct btrfs_raid_bio *rbio;
2129 int ret;
2130
Zhao Lei8e5cfb52015-01-20 15:11:33 +08002131 rbio = alloc_rbio(root, bbio, stripe_len);
Miao Xieaf8e2d12014-10-23 14:42:50 +08002132 if (IS_ERR(rbio)) {
Zhao Lei6e9606d2015-01-20 15:11:34 +08002133 if (generic_io)
2134 btrfs_put_bbio(bbio);
David Woodhouse53b381b2013-01-29 18:40:14 -05002135 return PTR_ERR(rbio);
Miao Xieaf8e2d12014-10-23 14:42:50 +08002136 }
David Woodhouse53b381b2013-01-29 18:40:14 -05002137
Miao Xie1b94b552014-11-06 16:14:21 +08002138 rbio->operation = BTRFS_RBIO_READ_REBUILD;
David Woodhouse53b381b2013-01-29 18:40:14 -05002139 bio_list_add(&rbio->bio_list, bio);
Kent Overstreet4f024f32013-10-11 15:44:27 -07002140 rbio->bio_list_bytes = bio->bi_iter.bi_size;
David Woodhouse53b381b2013-01-29 18:40:14 -05002141
2142 rbio->faila = find_logical_bio_stripe(rbio, bio);
2143 if (rbio->faila == -1) {
2144 BUG();
Zhao Lei6e9606d2015-01-20 15:11:34 +08002145 if (generic_io)
2146 btrfs_put_bbio(bbio);
David Woodhouse53b381b2013-01-29 18:40:14 -05002147 kfree(rbio);
2148 return -EIO;
2149 }
2150
Miao Xie42452152014-11-25 16:39:28 +08002151 if (generic_io) {
2152 btrfs_bio_counter_inc_noblocked(root->fs_info);
2153 rbio->generic_bio_cnt = 1;
2154 } else {
Zhao Lei6e9606d2015-01-20 15:11:34 +08002155 btrfs_get_bbio(bbio);
Miao Xie42452152014-11-25 16:39:28 +08002156 }
2157
David Woodhouse53b381b2013-01-29 18:40:14 -05002158 /*
2159 * reconstruct from the q stripe if they are
2160 * asking for mirror 3
2161 */
2162 if (mirror_num == 3)
Miao Xie2c8cdd62014-11-14 16:06:25 +08002163 rbio->failb = rbio->real_stripes - 2;
David Woodhouse53b381b2013-01-29 18:40:14 -05002164
2165 ret = lock_stripe_add(rbio);
2166
2167 /*
2168 * __raid56_parity_recover will end the bio with
2169 * any errors it hits. We don't want to return
2170 * its error value up the stack because our caller
2171 * will end up calling bio_endio with any nonzero
2172 * return
2173 */
2174 if (ret == 0)
2175 __raid56_parity_recover(rbio);
2176 /*
2177 * our rbio has been added to the list of
2178 * rbios that will be handled after the
2179 * currently lock owner is done
2180 */
2181 return 0;
2182
2183}
2184
2185static void rmw_work(struct btrfs_work *work)
2186{
2187 struct btrfs_raid_bio *rbio;
2188
2189 rbio = container_of(work, struct btrfs_raid_bio, work);
2190 raid56_rmw_stripe(rbio);
2191}
2192
2193static void read_rebuild_work(struct btrfs_work *work)
2194{
2195 struct btrfs_raid_bio *rbio;
2196
2197 rbio = container_of(work, struct btrfs_raid_bio, work);
2198 __raid56_parity_recover(rbio);
2199}
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002200
2201/*
2202 * The following code is used to scrub/replace the parity stripe
2203 *
2204 * Note: We need make sure all the pages that add into the scrub/replace
2205 * raid bio are correct and not be changed during the scrub/replace. That
2206 * is those pages just hold metadata or file data with checksum.
2207 */
2208
2209struct btrfs_raid_bio *
2210raid56_parity_alloc_scrub_rbio(struct btrfs_root *root, struct bio *bio,
Zhao Lei8e5cfb52015-01-20 15:11:33 +08002211 struct btrfs_bio *bbio, u64 stripe_len,
2212 struct btrfs_device *scrub_dev,
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002213 unsigned long *dbitmap, int stripe_nsectors)
2214{
2215 struct btrfs_raid_bio *rbio;
2216 int i;
2217
Zhao Lei8e5cfb52015-01-20 15:11:33 +08002218 rbio = alloc_rbio(root, bbio, stripe_len);
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002219 if (IS_ERR(rbio))
2220 return NULL;
2221 bio_list_add(&rbio->bio_list, bio);
2222 /*
2223 * This is a special bio which is used to hold the completion handler
2224 * and make the scrub rbio is similar to the other types
2225 */
2226 ASSERT(!bio->bi_iter.bi_size);
2227 rbio->operation = BTRFS_RBIO_PARITY_SCRUB;
2228
Miao Xie2c8cdd62014-11-14 16:06:25 +08002229 for (i = 0; i < rbio->real_stripes; i++) {
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002230 if (bbio->stripes[i].dev == scrub_dev) {
2231 rbio->scrubp = i;
2232 break;
2233 }
2234 }
2235
2236 /* Now we just support the sectorsize equals to page size */
2237 ASSERT(root->sectorsize == PAGE_SIZE);
2238 ASSERT(rbio->stripe_npages == stripe_nsectors);
2239 bitmap_copy(rbio->dbitmap, dbitmap, stripe_nsectors);
2240
2241 return rbio;
2242}
2243
Omar Sandovalb4ee1782015-06-19 11:52:50 -07002244/* Used for both parity scrub and missing. */
2245void raid56_add_scrub_pages(struct btrfs_raid_bio *rbio, struct page *page,
2246 u64 logical)
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002247{
2248 int stripe_offset;
2249 int index;
2250
Zhao Lei8e5cfb52015-01-20 15:11:33 +08002251 ASSERT(logical >= rbio->bbio->raid_map[0]);
2252 ASSERT(logical + PAGE_SIZE <= rbio->bbio->raid_map[0] +
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002253 rbio->stripe_len * rbio->nr_data);
Zhao Lei8e5cfb52015-01-20 15:11:33 +08002254 stripe_offset = (int)(logical - rbio->bbio->raid_map[0]);
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002255 index = stripe_offset >> PAGE_CACHE_SHIFT;
2256 rbio->bio_pages[index] = page;
2257}
2258
2259/*
2260 * We just scrub the parity that we have correct data on the same horizontal,
2261 * so we needn't allocate all pages for all the stripes.
2262 */
2263static int alloc_rbio_essential_pages(struct btrfs_raid_bio *rbio)
2264{
2265 int i;
2266 int bit;
2267 int index;
2268 struct page *page;
2269
2270 for_each_set_bit(bit, rbio->dbitmap, rbio->stripe_npages) {
Miao Xie2c8cdd62014-11-14 16:06:25 +08002271 for (i = 0; i < rbio->real_stripes; i++) {
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002272 index = i * rbio->stripe_npages + bit;
2273 if (rbio->stripe_pages[index])
2274 continue;
2275
2276 page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2277 if (!page)
2278 return -ENOMEM;
2279 rbio->stripe_pages[index] = page;
2280 ClearPageUptodate(page);
2281 }
2282 }
2283 return 0;
2284}
2285
2286/*
2287 * end io function used by finish_rmw. When we finally
2288 * get here, we've written a full stripe
2289 */
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02002290static void raid_write_parity_end_io(struct bio *bio)
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002291{
2292 struct btrfs_raid_bio *rbio = bio->bi_private;
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02002293 int err = bio->bi_error;
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002294
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02002295 if (bio->bi_error)
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002296 fail_bio_stripe(rbio, bio);
2297
2298 bio_put(bio);
2299
2300 if (!atomic_dec_and_test(&rbio->stripes_pending))
2301 return;
2302
2303 err = 0;
2304
2305 if (atomic_read(&rbio->error))
2306 err = -EIO;
2307
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02002308 rbio_orig_end_io(rbio, err);
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002309}
2310
2311static noinline void finish_parity_scrub(struct btrfs_raid_bio *rbio,
2312 int need_check)
2313{
Miao Xie76035972014-11-14 17:45:42 +08002314 struct btrfs_bio *bbio = rbio->bbio;
Miao Xie2c8cdd62014-11-14 16:06:25 +08002315 void *pointers[rbio->real_stripes];
Miao Xie76035972014-11-14 17:45:42 +08002316 DECLARE_BITMAP(pbitmap, rbio->stripe_npages);
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002317 int nr_data = rbio->nr_data;
2318 int stripe;
2319 int pagenr;
2320 int p_stripe = -1;
2321 int q_stripe = -1;
2322 struct page *p_page = NULL;
2323 struct page *q_page = NULL;
2324 struct bio_list bio_list;
2325 struct bio *bio;
Miao Xie76035972014-11-14 17:45:42 +08002326 int is_replace = 0;
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002327 int ret;
2328
2329 bio_list_init(&bio_list);
2330
Miao Xie2c8cdd62014-11-14 16:06:25 +08002331 if (rbio->real_stripes - rbio->nr_data == 1) {
2332 p_stripe = rbio->real_stripes - 1;
2333 } else if (rbio->real_stripes - rbio->nr_data == 2) {
2334 p_stripe = rbio->real_stripes - 2;
2335 q_stripe = rbio->real_stripes - 1;
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002336 } else {
2337 BUG();
2338 }
2339
Miao Xie76035972014-11-14 17:45:42 +08002340 if (bbio->num_tgtdevs && bbio->tgtdev_map[rbio->scrubp]) {
2341 is_replace = 1;
2342 bitmap_copy(pbitmap, rbio->dbitmap, rbio->stripe_npages);
2343 }
2344
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002345 /*
2346 * Because the higher layers(scrubber) are unlikely to
2347 * use this area of the disk again soon, so don't cache
2348 * it.
2349 */
2350 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
2351
2352 if (!need_check)
2353 goto writeback;
2354
2355 p_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2356 if (!p_page)
2357 goto cleanup;
2358 SetPageUptodate(p_page);
2359
2360 if (q_stripe != -1) {
2361 q_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2362 if (!q_page) {
2363 __free_page(p_page);
2364 goto cleanup;
2365 }
2366 SetPageUptodate(q_page);
2367 }
2368
2369 atomic_set(&rbio->error, 0);
2370
2371 for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) {
2372 struct page *p;
2373 void *parity;
2374 /* first collect one page from each data stripe */
2375 for (stripe = 0; stripe < nr_data; stripe++) {
2376 p = page_in_rbio(rbio, stripe, pagenr, 0);
2377 pointers[stripe] = kmap(p);
2378 }
2379
2380 /* then add the parity stripe */
2381 pointers[stripe++] = kmap(p_page);
2382
2383 if (q_stripe != -1) {
2384
2385 /*
2386 * raid6, add the qstripe and call the
2387 * library function to fill in our p/q
2388 */
2389 pointers[stripe++] = kmap(q_page);
2390
Miao Xie2c8cdd62014-11-14 16:06:25 +08002391 raid6_call.gen_syndrome(rbio->real_stripes, PAGE_SIZE,
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002392 pointers);
2393 } else {
2394 /* raid5 */
2395 memcpy(pointers[nr_data], pointers[0], PAGE_SIZE);
2396 run_xor(pointers + 1, nr_data - 1, PAGE_CACHE_SIZE);
2397 }
2398
2399 /* Check scrubbing pairty and repair it */
2400 p = rbio_stripe_page(rbio, rbio->scrubp, pagenr);
2401 parity = kmap(p);
2402 if (memcmp(parity, pointers[rbio->scrubp], PAGE_CACHE_SIZE))
2403 memcpy(parity, pointers[rbio->scrubp], PAGE_CACHE_SIZE);
2404 else
2405 /* Parity is right, needn't writeback */
2406 bitmap_clear(rbio->dbitmap, pagenr, 1);
2407 kunmap(p);
2408
Miao Xie2c8cdd62014-11-14 16:06:25 +08002409 for (stripe = 0; stripe < rbio->real_stripes; stripe++)
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002410 kunmap(page_in_rbio(rbio, stripe, pagenr, 0));
2411 }
2412
2413 __free_page(p_page);
2414 if (q_page)
2415 __free_page(q_page);
2416
2417writeback:
2418 /*
2419 * time to start writing. Make bios for everything from the
2420 * higher layers (the bio_list in our rbio) and our p/q. Ignore
2421 * everything else.
2422 */
2423 for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) {
2424 struct page *page;
2425
2426 page = rbio_stripe_page(rbio, rbio->scrubp, pagenr);
2427 ret = rbio_add_io_page(rbio, &bio_list,
2428 page, rbio->scrubp, pagenr, rbio->stripe_len);
2429 if (ret)
2430 goto cleanup;
2431 }
2432
Miao Xie76035972014-11-14 17:45:42 +08002433 if (!is_replace)
2434 goto submit_write;
2435
2436 for_each_set_bit(pagenr, pbitmap, rbio->stripe_npages) {
2437 struct page *page;
2438
2439 page = rbio_stripe_page(rbio, rbio->scrubp, pagenr);
2440 ret = rbio_add_io_page(rbio, &bio_list, page,
2441 bbio->tgtdev_map[rbio->scrubp],
2442 pagenr, rbio->stripe_len);
2443 if (ret)
2444 goto cleanup;
2445 }
2446
2447submit_write:
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002448 nr_data = bio_list_size(&bio_list);
2449 if (!nr_data) {
2450 /* Every parity is right */
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02002451 rbio_orig_end_io(rbio, 0);
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002452 return;
2453 }
2454
2455 atomic_set(&rbio->stripes_pending, nr_data);
2456
2457 while (1) {
2458 bio = bio_list_pop(&bio_list);
2459 if (!bio)
2460 break;
2461
2462 bio->bi_private = rbio;
2463 bio->bi_end_io = raid_write_parity_end_io;
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002464 submit_bio(WRITE, bio);
2465 }
2466 return;
2467
2468cleanup:
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02002469 rbio_orig_end_io(rbio, -EIO);
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002470}
2471
2472static inline int is_data_stripe(struct btrfs_raid_bio *rbio, int stripe)
2473{
2474 if (stripe >= 0 && stripe < rbio->nr_data)
2475 return 1;
2476 return 0;
2477}
2478
2479/*
2480 * While we're doing the parity check and repair, we could have errors
2481 * in reading pages off the disk. This checks for errors and if we're
2482 * not able to read the page it'll trigger parity reconstruction. The
2483 * parity scrub will be finished after we've reconstructed the failed
2484 * stripes
2485 */
2486static void validate_rbio_for_parity_scrub(struct btrfs_raid_bio *rbio)
2487{
2488 if (atomic_read(&rbio->error) > rbio->bbio->max_errors)
2489 goto cleanup;
2490
2491 if (rbio->faila >= 0 || rbio->failb >= 0) {
2492 int dfail = 0, failp = -1;
2493
2494 if (is_data_stripe(rbio, rbio->faila))
2495 dfail++;
2496 else if (is_parity_stripe(rbio->faila))
2497 failp = rbio->faila;
2498
2499 if (is_data_stripe(rbio, rbio->failb))
2500 dfail++;
2501 else if (is_parity_stripe(rbio->failb))
2502 failp = rbio->failb;
2503
2504 /*
2505 * Because we can not use a scrubbing parity to repair
2506 * the data, so the capability of the repair is declined.
2507 * (In the case of RAID5, we can not repair anything)
2508 */
2509 if (dfail > rbio->bbio->max_errors - 1)
2510 goto cleanup;
2511
2512 /*
2513 * If all data is good, only parity is correctly, just
2514 * repair the parity.
2515 */
2516 if (dfail == 0) {
2517 finish_parity_scrub(rbio, 0);
2518 return;
2519 }
2520
2521 /*
2522 * Here means we got one corrupted data stripe and one
2523 * corrupted parity on RAID6, if the corrupted parity
2524 * is scrubbing parity, luckly, use the other one to repair
2525 * the data, or we can not repair the data stripe.
2526 */
2527 if (failp != rbio->scrubp)
2528 goto cleanup;
2529
2530 __raid_recover_end_io(rbio);
2531 } else {
2532 finish_parity_scrub(rbio, 1);
2533 }
2534 return;
2535
2536cleanup:
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02002537 rbio_orig_end_io(rbio, -EIO);
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002538}
2539
2540/*
2541 * end io for the read phase of the rmw cycle. All the bios here are physical
2542 * stripe bios we've read from the disk so we can recalculate the parity of the
2543 * stripe.
2544 *
2545 * This will usually kick off finish_rmw once all the bios are read in, but it
2546 * may trigger parity reconstruction if we had any errors along the way
2547 */
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02002548static void raid56_parity_scrub_end_io(struct bio *bio)
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002549{
2550 struct btrfs_raid_bio *rbio = bio->bi_private;
2551
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02002552 if (bio->bi_error)
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002553 fail_bio_stripe(rbio, bio);
2554 else
2555 set_bio_pages_uptodate(bio);
2556
2557 bio_put(bio);
2558
2559 if (!atomic_dec_and_test(&rbio->stripes_pending))
2560 return;
2561
2562 /*
2563 * this will normally call finish_rmw to start our write
2564 * but if there are any failed stripes we'll reconstruct
2565 * from parity first
2566 */
2567 validate_rbio_for_parity_scrub(rbio);
2568}
2569
2570static void raid56_parity_scrub_stripe(struct btrfs_raid_bio *rbio)
2571{
2572 int bios_to_read = 0;
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002573 struct bio_list bio_list;
2574 int ret;
2575 int pagenr;
2576 int stripe;
2577 struct bio *bio;
2578
2579 ret = alloc_rbio_essential_pages(rbio);
2580 if (ret)
2581 goto cleanup;
2582
2583 bio_list_init(&bio_list);
2584
2585 atomic_set(&rbio->error, 0);
2586 /*
2587 * build a list of bios to read all the missing parts of this
2588 * stripe
2589 */
Miao Xie2c8cdd62014-11-14 16:06:25 +08002590 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002591 for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) {
2592 struct page *page;
2593 /*
2594 * we want to find all the pages missing from
2595 * the rbio and read them from the disk. If
2596 * page_in_rbio finds a page in the bio list
2597 * we don't need to read it off the stripe.
2598 */
2599 page = page_in_rbio(rbio, stripe, pagenr, 1);
2600 if (page)
2601 continue;
2602
2603 page = rbio_stripe_page(rbio, stripe, pagenr);
2604 /*
2605 * the bio cache may have handed us an uptodate
2606 * page. If so, be happy and use it
2607 */
2608 if (PageUptodate(page))
2609 continue;
2610
2611 ret = rbio_add_io_page(rbio, &bio_list, page,
2612 stripe, pagenr, rbio->stripe_len);
2613 if (ret)
2614 goto cleanup;
2615 }
2616 }
2617
2618 bios_to_read = bio_list_size(&bio_list);
2619 if (!bios_to_read) {
2620 /*
2621 * this can happen if others have merged with
2622 * us, it means there is nothing left to read.
2623 * But if there are missing devices it may not be
2624 * safe to do the full stripe write yet.
2625 */
2626 goto finish;
2627 }
2628
2629 /*
2630 * the bbio may be freed once we submit the last bio. Make sure
2631 * not to touch it after that
2632 */
2633 atomic_set(&rbio->stripes_pending, bios_to_read);
2634 while (1) {
2635 bio = bio_list_pop(&bio_list);
2636 if (!bio)
2637 break;
2638
2639 bio->bi_private = rbio;
2640 bio->bi_end_io = raid56_parity_scrub_end_io;
2641
2642 btrfs_bio_wq_end_io(rbio->fs_info, bio,
2643 BTRFS_WQ_ENDIO_RAID56);
2644
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002645 submit_bio(READ, bio);
2646 }
2647 /* the actual write will happen once the reads are done */
2648 return;
2649
2650cleanup:
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02002651 rbio_orig_end_io(rbio, -EIO);
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002652 return;
2653
2654finish:
2655 validate_rbio_for_parity_scrub(rbio);
2656}
2657
2658static void scrub_parity_work(struct btrfs_work *work)
2659{
2660 struct btrfs_raid_bio *rbio;
2661
2662 rbio = container_of(work, struct btrfs_raid_bio, work);
2663 raid56_parity_scrub_stripe(rbio);
2664}
2665
2666static void async_scrub_parity(struct btrfs_raid_bio *rbio)
2667{
2668 btrfs_init_work(&rbio->work, btrfs_rmw_helper,
2669 scrub_parity_work, NULL, NULL);
2670
2671 btrfs_queue_work(rbio->fs_info->rmw_workers,
2672 &rbio->work);
2673}
2674
2675void raid56_parity_submit_scrub_rbio(struct btrfs_raid_bio *rbio)
2676{
2677 if (!lock_stripe_add(rbio))
2678 async_scrub_parity(rbio);
2679}
Omar Sandovalb4ee1782015-06-19 11:52:50 -07002680
2681/* The following code is used for dev replace of a missing RAID 5/6 device. */
2682
2683struct btrfs_raid_bio *
2684raid56_alloc_missing_rbio(struct btrfs_root *root, struct bio *bio,
2685 struct btrfs_bio *bbio, u64 length)
2686{
2687 struct btrfs_raid_bio *rbio;
2688
2689 rbio = alloc_rbio(root, bbio, length);
2690 if (IS_ERR(rbio))
2691 return NULL;
2692
2693 rbio->operation = BTRFS_RBIO_REBUILD_MISSING;
2694 bio_list_add(&rbio->bio_list, bio);
2695 /*
2696 * This is a special bio which is used to hold the completion handler
2697 * and make the scrub rbio is similar to the other types
2698 */
2699 ASSERT(!bio->bi_iter.bi_size);
2700
2701 rbio->faila = find_logical_bio_stripe(rbio, bio);
2702 if (rbio->faila == -1) {
2703 BUG();
2704 kfree(rbio);
2705 return NULL;
2706 }
2707
2708 return rbio;
2709}
2710
2711static void missing_raid56_work(struct btrfs_work *work)
2712{
2713 struct btrfs_raid_bio *rbio;
2714
2715 rbio = container_of(work, struct btrfs_raid_bio, work);
2716 __raid56_parity_recover(rbio);
2717}
2718
2719static void async_missing_raid56(struct btrfs_raid_bio *rbio)
2720{
2721 btrfs_init_work(&rbio->work, btrfs_rmw_helper,
2722 missing_raid56_work, NULL, NULL);
2723
2724 btrfs_queue_work(rbio->fs_info->rmw_workers, &rbio->work);
2725}
2726
2727void raid56_submit_missing_rbio(struct btrfs_raid_bio *rbio)
2728{
2729 if (!lock_stripe_add(rbio))
2730 async_missing_raid56(rbio);
2731}